Introduction
The Standard Model (SM) of particle physics, completed by the discovery of the Higgs boson [1 , 2 ] , describes all observed collider phenomena with remarkable accuracy and remains consistent with essentially every precision test performed to date.
Despite this success, the SM cannot be the complete theory of nature. It does not account for neutrino masses, dark matter, or the origin of the observed cosmic matter–antimatter asymmetry, leaving several central problems of cosmology and particle physics unresolved.
Among these, the baryon asymmetry of the universe (BAU) is a particularly sharp challenge: the SM cannot generate the observed asymmetry through the Sakharov conditions in sufficient quantity [3 , 4 , 5 , 6 ] . This motivates extending the scalar sector of the SM in ways that can support a viable baryogenesis mechanism while remaining compatible with current collider and dark matter constraints.
Electroweak baryogenesis (EWBG) offers an attractive mechanism to generate the BAU at the electroweak scale, but it requires two separate ingredients absent from the SM: a strongly first-order EWPT and new sources of CP violation. In the SM itself neither is realized – lattice studies show that the electroweak transition is a smooth crossover rather than a first-order transition [7 ] , and the CP violation of the CKM matrix is far too small to seed the observed asymmetry – so EWBG is only a viable option once the scalar sector is extended. Extending the SM with an additional scalar, most simply a real singlet, is one of the best-studied ways to restore a strongly first-order EWPT: the extra scalar can modify the shape of the finite-temperature potential enough to generate a barrier between the symmetric and broken phases, and such singlet extensions have been extensively analyzed in connection with dark matter, vacuum stability, and the EWPT, both for the real, Z 2 Z_2 Z 2 -symmetric case that we adopt below [8 , 9 , 10 , 11 , 12 ] and, once the singlet is complexified or coupled with additional CP-odd interactions, as a source of the CP violation needed to complete EWBG [13 , 14 ] . Reliable computations of the finite-temperature effective potential are essential for establishing whether a given extension actually achieves a strongly first-order transition, but conventional perturbative techniques suffer from severe IR sensitivity and resummation issues near the critical temperature of a first-order EWPT.
The functional renormalization group (FRG), formulated via the Wetterich equation, provides a non-perturbative tool which smoothly interpolates between the bare UV action and the full IR effective action, and is well suited to resolving these IR issues. However, in scalar extensions of the SM the effective potential depends on multiple field invariants, leading to a nonlinear PDE in more than one variable, which is technically demanding to solve by standard finite-difference or pseudo-spectral methods. Most functional renormalization group studies of scalar extensions of the Standard Model rely on low-order polynomial truncations of the effective potential or grid-based discretizations in field space.
Here we consider the real singlet scalar extension of the SM with Z 2 Z_2 Z 2 symmetry as a concrete setting for this FRG treatment. Being Z 2 Z_2 Z 2 -symmetric and real, the singlet introduces no new source of CP violation, so the model by itself cannot complete EWBG; we take the strongly first-order transition as the first step to be established quantitatively, using this model as a controlled testing ground for the finite-temperature effective-potential methodology, and leave the additional CP-violating sector required to generate the BAU to future extensions of the model.
Physics-informed neural networks (PINNs) [15 ] parametrize the unknown function as a neural network and minimize the PDE residual via automatic differentiation. They are mesh-free and naturally handle multi-dimensional domains, making them a natural candidate for the multi-field FRG flow equations that are otherwise costly to solve. Applying this idea to the FRG flow equation itself is a recent development: a genuinely mesh-free treatment, in which a neural network represents the field dependence continuously and the flow equation is enforced through automatic differentiation of the network output, has been demonstrated for a single-field O ( N ) O(N) O ( N ) -symmetric scalar theory [16 ] , and a related neural-network parametrization of the field- and scale-dependent potential for the O ( N ) O(N) O ( N ) model at finite temperature has also been proposed, though there the network is trained against a finite-difference-discretized residual evaluated on a fixed field-and-RG-time grid rather than a continuous PDE residual [17 ] . Extending a genuinely continuous, autodiff-based PINN treatment to a multi-field, finite-temperature effective potential of a phenomenologically motivated BSM extension – as required for a first-order EWPT analysis – remains unexplored. In this paper we apply PINNs to the FRG equation of the Z 2 Z_2 Z 2 -symmetric singlet extension of the SM with finite-temperature contributions and SM gauge/Yukawa loops. Our aim here is methodological: to establish the end-to-end pipeline – from UV couplings, through the mesh-free flow, to a finite-temperature effective potential and a bounce action – as a proof of concept, rather than to deliver a full temperature scan or a converged determination of the transition strength at this benchmark.
The remainder of this paper is organized as follows. Sec. 2 sets up the Wetterich equation for the Z 2 \mathbb{Z}_2 Z 2 singlet extension and its dimensionless LPA′ ' ′ truncation. Sec. 3 reviews the resulting two-step electroweak phase transition and the critical and nucleation temperatures used to assess a strongly first-order transition. Sec. 4 fixes the benchmark point and input parameters used throughout, including the one-loop Coleman–Weinberg matching at k I R k_{\mathrm{IR}} k IR and the two-loop RGE running to k U V k_{\mathrm{UV}} k UV , and introduces a grid-based Newton–Krylov relaxation solver as a benchmark for the flow equation. It first checks the solver on a singlet-free reduction of the model – the simplest, single-field version of the problem – and then on the full two-field system, so as to test whether the observed convergence failure originates in the Higgs–singlet mass-eigenvalue mixing or is already present in the single-field loop-level flow equation. Sec. 5 describes the PINN architecture, decomposition of the potential, loss function, and multi-block training scheme used to solve the same flow equation. Sec. 6 presents the resulting PINN solutions, again starting from the singlet-free reduction before the full system, compares them with the perturbative and relaxation-based benchmarks, and discusses the practical role and limitations of the soft perturbative-consistency penalty used in the PINN loss. Sec. 7 summarizes our findings and outlines directions for future work, including the CP-violating extensions needed to complete EWBG.
Wetterich equation for the Z 2 \mathbb{Z}_2 Z 2 singlet extension
Model and truncation
We consider the scalar sector of the Standard Model extended by a real singlet scalar S S S with a Z 2 \mathbb{Z}_2 Z 2 symmetry,
S → − S . (1) S \to-S. \tag*{(1)} S → − S . (1)
In this subsection, we keep only scalar fluctuations and neglect gauge and Yukawa contributions. At the UV scale k = Λ k = \Lambda k = Λ , the Z 2 \mathbb{Z}_2 Z 2 -symmetric scalar potential is
V Λ ( H , S ) = μ H 2 ∣ H ∣ 2 + λ H ∣ H ∣ 4 + 1 2 μ S 2 S 2 + 1 4 λ S S 4 + 1 2 λ H S ∣ H ∣ 2 S 2 . (2) V_{\Lambda}(H,S) = \mu_H^2 |H|^2 + \lambda_H |H|^4 + \frac{1}{2}\mu_S^2 S^2 + \frac{1}{4}\lambda_S S^4 + \frac{1}{2}\lambda_{HS}|H|^2S^2. \tag*{(2)} V Λ ( H , S ) = μ H 2 ∣ H ∣ 2 + λ H ∣ H ∣ 4 + 2 1 μ S 2 S 2 + 4 1 λ S S 4 + 2 1 λ H S ∣ H ∣ 2 S 2 . (2)
Introducing the field invariants
ρ ~ ≡ H † H , σ ~ ≡ 1 2 S 2 , (3) \tilde{\rho} \equiv H^\dagger H,\qquad\tilde{\sigma} \equiv\frac{1}{2}S^2, \tag*{(3)} ρ ~ ≡ H † H , σ ~ ≡ 2 1 S 2 , (3)
the potential can be written as
V Λ ( ρ ~ , σ ~ ) = μ H 2 ρ ~ + λ H ρ ~ 2 + μ S 2 σ ~ + λ S σ ~ 2 + λ H S ρ ~ σ ~ . (4) V_{\Lambda}(\tilde{\rho},\tilde{\sigma}) = \mu_H^2\tilde{\rho} + \lambda_H\tilde{\rho}^2 + \mu_S^2\tilde{\sigma} + \lambda_S\tilde{\sigma}^2 + \lambda_{HS}\tilde{\rho}\tilde{\sigma}. \tag*{(4)} V Λ ( ρ ~ , σ ~ ) = μ H 2 ρ ~ + λ H ρ ~ 2 + μ S 2 σ ~ + λ S σ ~ 2 + λ H S ρ ~ σ ~ . (4)
Within the derivative expansion at LPA′ ' ′ [18 ] , we employ the ansatz
Γ k [ H , S ] = ∫ x [ Z H , k ( ∂ μ H ) † ( ∂ μ H ) + 1 2 Z S , k ( ∂ μ S ) 2 + V k ( ρ ~ , σ ~ ) ] . (5) \Gamma_k[H,S] = \int_x \left[ Z_{H,k}(\partial_\mu H)^\dagger(\partial_\mu H) + \frac{1}{2}Z_{S,k}(\partial_\mu S)^2 + V_k(\tilde{\rho},\tilde{\sigma}) \right]. \tag*{(5)} Γ k [ H , S ] = ∫ x [ Z H , k ( ∂ μ H ) † ( ∂ μ H ) + 2 1 Z S , k ( ∂ μ S ) 2 + V k ( ρ ~ , σ ~ ) ] . (5)
where
∫ x ≡ ∫ 0 1 / T d τ ∫ d 3 x , t ≡ ln k Λ . (6) \int_x \equiv\int_0^{1/T} d\tau\int d^3x,\qquad t \equiv\ln\frac{k}{\Lambda}. \tag*{(6)} ∫ x ≡ ∫ 0 1/ T d τ ∫ d 3 x , t ≡ ln Λ k . (6)
and the anomalous dimensions are defined by
η H ≡ − ∂ t ln Z H , k , η S ≡ − ∂ t ln Z S , k . (7) \eta_H \equiv-\partial_t \ln Z_{H,k},\qquad\eta_S \equiv-\partial_t \ln Z_{S,k}. \tag*{(7)} η H ≡ − ∂ t ln Z H , k , η S ≡ − ∂ t ln Z S , k . (7)
Wetterich equation
The flow of the effective average action is governed by the Wetterich equation [19 , 20 ] ,
∂ t Γ k = 1 2 STr [ ( Γ k ( 2 ) + R k ) − 1 ∂ t R k ] . (8) \partial_t\Gamma_k = \frac{1}{2}\operatorname{STr}\left[\left(\Gamma_k^{(2)}+R_k\right)^{-1}\partial_tR_k\right]. \tag*{(8)} ∂ t Γ k = 2 1 STr [ ( Γ k ( 2 ) + R k ) − 1 ∂ t R k ] . (8)
At finite temperature, the Euclidean zeroth component is replaced by the bosonic Matsubara frequencies,
p 0 → ω n = 2 π n T , ∫ d p 0 2 π → T ∑ n ∈ Z . (9) p_0 \to\omega_n = 2\pi nT,\qquad\int\frac{\mathrm{d}p_0}{2\pi}\to T\sum_{n\in\mathbb{Z}}. \tag*{(9)} p 0 → ω n = 2 π n T , ∫ 2 π d p 0 → T n ∈ Z ∑ . (9)
Using the Litim regulator [21 ] ,
R H , k ( p ) = Z H , k ( k 2 − p 2 ) θ ( k 2 − p 2 ) , (10) R_{H,k}(p)=Z_{H,k}\left(k^2-p^2\right)\theta\left(k^2-p^2\right), \tag*{(10)} R H , k ( p ) = Z H , k ( k 2 − p 2 ) θ ( k 2 − p 2 ) , (10)
R S , k ( p ) = Z S , k ( k 2 − p 2 ) θ ( k 2 − p 2 ) , (11) R_{S,k}(p)=Z_{S,k}\left(k^2-p^2\right)\theta\left(k^2-p^2\right), \tag*{(11)} R S , k ( p ) = Z S , k ( k 2 − p 2 ) θ ( k 2 − p 2 ) , (11)
and evaluating the flow for homogeneous background fields, the momentum-shell integral is elementary and the Matsubara sum closes in terms of a coth [22 ] , giving the flow of the effective potential,
∂ t V k ( ρ ~ , σ ~ ) = k 4 12 π 2 [ 3 1 + m ˉ G 2 ( ρ ~ , σ ~ ) coth ( 1 + m ˉ G 2 ( ρ ~ , σ ~ ) 2 τ ) + 1 1 + m ˉ + 2 ( ρ ~ , σ ~ ) coth ( 1 + m ˉ + 2 ( ρ ~ , σ ~ ) 2 τ ) + 1 1 + m ˉ − 2 ( ρ ~ , σ ~ ) coth ( 1 + m ˉ − 2 ( ρ ~ , σ ~ ) 2 τ ) ] + k 4 E e x t , (12) \begin{aligned}
\partial_tV_k(\tilde{\rho},\tilde{\sigma})={}&\frac{k^4}{12\pi^2}\left[\frac{3}{\sqrt{1+\bar{m}_G^2(\tilde{\rho},\tilde{\sigma})}}\coth\left(\frac{\sqrt{1+\bar{m}_G^2(\tilde{\rho},\tilde{\sigma})}}{2\tau}\right)\right.\\
&+\frac{1}{\sqrt{1+\bar{m}_+^2(\tilde{\rho},\tilde{\sigma})}}\coth\left(\frac{\sqrt{1+\bar{m}_+^2(\tilde{\rho},\tilde{\sigma})}}{2\tau}\right)\\
&\left.+\frac{1}{\sqrt{1+\bar{m}_-^2(\tilde{\rho},\tilde{\sigma})}}\coth\left(\frac{\sqrt{1+\bar{m}_-^2(\tilde{\rho},\tilde{\sigma})}}{2\tau}\right)\right]+k^4E_{\mathrm{ext}}, \tag*{(12)}
\end{aligned} ∂ t V k ( ρ ~ , σ ~ ) = 12 π 2 k 4 [ 1 + m ˉ G 2 ( ρ ~ , σ ~ ) 3 coth ( 2 τ 1 + m ˉ G 2 ( ρ ~ , σ ~ ) ) + 1 + m ˉ + 2 ( ρ ~ , σ ~ ) 1 coth ( 2 τ 1 + m ˉ + 2 ( ρ ~ , σ ~ ) ) + 1 + m ˉ − 2 ( ρ ~ , σ ~ ) 1 coth ( 2 τ 1 + m ˉ − 2 ( ρ ~ , σ ~ ) ) ] + k 4 E ext , (12)
where τ = T / k \tau=T/k τ = T / k . Each bosonic mode i i i in the bracket arises as
1 2 ∫ d 3 p ( 2 π ) 3 T ∑ n ∈ Z ∂ t R k ω n 2 + p 2 + R k + m ˉ i 2 k 2 = k 4 12 π 2 1 1 + m ˉ i 2 coth ( k 1 + m ˉ i 2 2 T ) , (13) \frac{1}{2}\int\frac{\mathrm{d}^3p}{(2\pi)^3}T\sum_{n\in\mathbb{Z}}\frac{\partial_tR_k}{\omega_n^2+p^2+R_k+\bar{m}_i^2k^2}=\frac{k^4}{12\pi^2}\frac{1}{\sqrt{1+\bar{m}_i^2}}\coth\left(\frac{k\sqrt{1+\bar{m}_i^2}}{2T}\right), \tag*{(13)} 2 1 ∫ ( 2 π ) 3 d 3 p T n ∈ Z ∑ ω n 2 + p 2 + R k + m ˉ i 2 k 2 ∂ t R k = 12 π 2 k 4 1 + m ˉ i 2 1 coth ( 2 T k 1 + m ˉ i 2 ) , (13)
since the θ \theta θ -function collapses the momentum-shell integral to ∫ ∣ p ∣ < k d 3 p / ( 2 π ) 3 = k 3 / 6 π 2 \int_{\lvert p\rvert<k}\mathrm{d}^3p/(2\pi)^3=k^3/6\pi^2 ∫ ∣ p ∣ < k d 3 p / ( 2 π ) 3 = k 3 /6 π 2 with ∂ t R k → 2 k 2 \partial_tR_k\to2k^2 ∂ t R k → 2 k 2 inside it, while the Matsubara sum T ∑ n ( ω n 2 + E i 2 ) − 1 = ( 2 E i ) − 1 coth ( E i / 2 T ) T\sum_n(\omega_n^2+E_i^2)^{-1}=(2E_i)^{-1}\coth(E_i/2T) T ∑ n ( ω n 2 + E i 2 ) − 1 = ( 2 E i ) − 1 coth ( E i /2 T ) , with E i = k 1 + m ˉ i 2 E_i=k\sqrt{1+\bar{m}_i^2} E i = k 1 + m ˉ i 2 , is elementary; the bracket sums this over the three Goldstone modes and the two radial eigenvalues. The term E e x t E_{\mathrm{ext}} E ext is the dimensionless sum of contributions from top quark and electroweak gauge boson loops (W and Z), evaluated within the same regulator scheme; it is given explicitly in Eq. (19) below and enters the dimensionless flow (17) directly, without the k 4 k^4 k 4 prefactor. Here, the three Goldstone modes contribute through
m ˉ G 2 ( ρ ~ , σ ~ ) = 1 k 2 Z H , k ∂ V k ∂ ρ ~ , (14) \bar{m}_G^2(\tilde{\rho},\tilde{\sigma})=\frac{1}{k^2Z_{H,k}}\frac{\partial V_k}{\partial\tilde{\rho}}, \tag*{(14)} m ˉ G 2 ( ρ ~ , σ ~ ) = k 2 Z H , k 1 ∂ ρ ~ ∂ V k , (14)
while m ˉ ± 2 \bar{m}_{\pm}^2 m ˉ ± 2 are the eigenvalues of the 2 × 2 2\times2 2 × 2 scalar radial mass matrix M ˉ 2 \bar{M}^2 M ˉ 2 in the Higgs–singlet sector. Their dimensionless forms are given in § 2.3 .
Dimensionless form
We introduce the dimensionless renormalized variables
u ( ρ , σ ) ≡ V k ( ρ ~ , σ ~ ) k 4 , ρ ≡ Z H , k ρ ~ k 2 , σ ≡ Z S , k σ ~ k 2 , (15) u(\rho,\sigma) \equiv\frac{V_k(\tilde{\rho},\tilde{\sigma})}{k^4}, \qquad\rho\equiv\frac{Z_{H,k}\tilde{\rho}}{k^2}, \qquad\sigma\equiv\frac{Z_{S,k}\tilde{\sigma}}{k^2}, \tag*{(15)} u ( ρ , σ ) ≡ k 4 V k ( ρ ~ , σ ~ ) , ρ ≡ k 2 Z H , k ρ ~ , σ ≡ k 2 Z S , k σ ~ , (15)
where the field invariants are normalized as ρ ~ = H † H = h 2 / 2 \tilde{\rho}=H^\dagger H=h^2/2 ρ ~ = H † H = h 2 /2 and σ ~ = s 2 / 2 \tilde{\sigma}=s^2/2 σ ~ = s 2 /2 , with h h h , s s s the physical (canonically normalized) radial fields and ⟨ h ⟩ = v = 246.2 G e V \langle h\rangle=v=246.2\,\mathrm{GeV} ⟨ h ⟩ = v = 246.2 GeV at the electroweak minimum. The flow equation and the neural-network representation work throughout with the dimensionless ( ρ , σ ) (\rho,\sigma) ( ρ , σ ) ; the physical field-space quantities of § 3 (vacuum locations, v c v_c v c , and the bounce trajectory ϕ = ( h , s ) \phi=(h,s) ϕ = ( h , s ) ) are quoted in the dimensionful invariants ( ρ ~ , σ ~ ) (\tilde{\rho},\tilde{\sigma}) ( ρ ~ , σ ~ ) in G e V 2 \mathrm{GeV}^2 GeV 2 . The UV potential becomes
u Λ ( ρ , σ ) = a H ρ + λ H ρ 2 + a S σ + λ S σ 2 + λ H S ρ σ , (16) u_\Lambda(\rho,\sigma)=a_H\rho+\lambda_H\rho^2+a_S\sigma+\lambda_S\sigma^2+\lambda_{HS}\rho\sigma, \tag*{(16)} u Λ ( ρ , σ ) = a H ρ + λ H ρ 2 + a S σ + λ S σ 2 + λ H S ρ σ , (16)
where a H = μ H 2 / Λ 2 a_H=\mu_H^2/\Lambda^2 a H = μ H 2 / Λ 2 and a S = μ S 2 / Λ 2 a_S=\mu_S^2/\Lambda^2 a S = μ S 2 / Λ 2 . Then the flow equation becomes
∂ t u ( ρ , σ ) = − 4 u ( ρ , σ ) + ( 2 + η H ) ρ u ρ + ( 2 + η S ) σ u σ + 1 12 π 2 [ 3 1 + m ˉ G 2 ( ρ , σ ) coth ( 1 + m ˉ G 2 ( ρ , σ ) 2 τ ) + 1 1 + m ˉ + 2 ( ρ , σ ) coth ( 1 + m ˉ + 2 ( ρ , σ ) 2 τ ) + 1 1 + m ˉ − 2 ( ρ , σ ) coth ( 1 + m ˉ − 2 ( ρ , σ ) 2 τ ) ] + E e x t ( t , τ , ρ ) (17) \begin{aligned}
\partial_t u(\rho,\sigma)={}&-4u(\rho,\sigma)+(2+\eta_H)\rho\,u_\rho+(2+\eta_S)\sigma\,u_\sigma\\
&+\frac{1}{12\pi^2}\left[\frac{3}{\sqrt{1+\bar{m}_G^2(\rho,\sigma)}}\coth\left(\frac{\sqrt{1+\bar{m}_G^2(\rho,\sigma)}}{2\tau}\right)\right.\\
&\left.+\frac{1}{\sqrt{1+\bar{m}_+^2(\rho,\sigma)}}\coth\left(\frac{\sqrt{1+\bar{m}_+^2(\rho,\sigma)}}{2\tau}\right)\right.\\
&\left.+\frac{1}{\sqrt{1+\bar{m}_-^2(\rho,\sigma)}}\coth\left(\frac{\sqrt{1+\bar{m}_-^2(\rho,\sigma)}}{2\tau}\right)\right]\\
&+E_{\mathrm{ext}}(t,\tau,\rho) \tag*{(17)}
\end{aligned} ∂ t u ( ρ , σ ) = − 4 u ( ρ , σ ) + ( 2 + η H ) ρ u ρ + ( 2 + η S ) σ u σ + 12 π 2 1 [ 1 + m ˉ G 2 ( ρ , σ ) 3 coth ( 2 τ 1 + m ˉ G 2 ( ρ , σ ) ) + 1 + m ˉ + 2 ( ρ , σ ) 1 coth ( 2 τ 1 + m ˉ + 2 ( ρ , σ ) ) + 1 + m ˉ − 2 ( ρ , σ ) 1 coth ( 2 τ 1 + m ˉ − 2 ( ρ , σ ) ) ] + E ext ( t , τ , ρ ) (17)
where
u ρ ≡ ∂ u ∂ ρ , u σ ≡ ∂ u ∂ σ , u ρ ρ ≡ ∂ 2 u ∂ ρ 2 , u σ σ ≡ ∂ 2 u ∂ σ 2 , u ρ σ ≡ ∂ 2 u ∂ ρ ∂ σ . (18) u_\rho\equiv\frac{\partial u}{\partial\rho},\qquad u_\sigma\equiv\frac{\partial u}{\partial\sigma},\qquad u_{\rho\rho}\equiv\frac{\partial^2u}{\partial\rho^2},\qquad u_{\sigma\sigma}\equiv\frac{\partial^2u}{\partial\sigma^2},\qquad u_{\rho\sigma}\equiv\frac{\partial^2u}{\partial\rho\partial\sigma}. \tag*{(18)} u ρ ≡ ∂ ρ ∂ u , u σ ≡ ∂ σ ∂ u , u ρρ ≡ ∂ ρ 2 ∂ 2 u , u σ σ ≡ ∂ σ 2 ∂ 2 u , u ρ σ ≡ ∂ ρ ∂ σ ∂ 2 u . (18)
In the zero-temperature limit, τ → 0 \tau\to0 τ → 0 , one has coth ( x / 2 τ ) → 1 \coth(x/2\tau)\to1 coth ( x /2 τ ) → 1 , and the above equation reduces to the usual T = 0 T=0 T = 0 flow.
Explicitly,
E e x t ( t , τ , ρ ) = 1 12 π 2 [ 6 1 + m ˉ W 2 coth ( 1 + m ˉ W 2 2 τ ) + 3 1 + m ˉ Z 2 coth ( 1 + m ˉ Z 2 2 τ ) − 12 1 + m ˉ t 2 tanh ( 1 + m ˉ t 2 2 τ ) ] , (19) E_{\mathrm{ext}}(t,\tau,\rho)=\frac{1}{12\pi^2}\left[\frac{6}{\sqrt{1+\bar{m}_W^2}}\coth\left(\frac{\sqrt{1+\bar{m}_W^2}}{2\tau}\right)+\frac{3}{\sqrt{1+\bar{m}_Z^2}}\coth\left(\frac{\sqrt{1+\bar{m}_Z^2}}{2\tau}\right)-\frac{12}{\sqrt{1+\bar{m}_t^2}}\tanh\left(\frac{\sqrt{1+\bar{m}_t^2}}{2\tau}\right)\right], \tag*{(19)} E ext ( t , τ , ρ ) = 12 π 2 1 [ 1 + m ˉ W 2 6 coth ( 2 τ 1 + m ˉ W 2 ) + 1 + m ˉ Z 2 3 coth ( 2 τ 1 + m ˉ Z 2 ) − 1 + m ˉ t 2 12 tanh ( 2 τ 1 + m ˉ t 2 ) ] , (19)
with dimensionless tree-level masses m ˉ W 2 = g 2 2 ρ / 2 \bar{m}_W^2=g_2^2\rho/2 m ˉ W 2 = g 2 2 ρ /2 , m ˉ Z 2 = ( g 1 2 + g 2 2 ) ρ / 2 \bar{m}_Z^2=(g_1^2+g_2^2)\rho/2 m ˉ Z 2 = ( g 1 2 + g 2 2 ) ρ /2 , and m ˉ t 2 = y t 2 ρ \bar{m}_t^2=y_t^2\rho m ˉ t 2 = y t 2 ρ (evaluated with the running g 1 g_1 g 1 , g 2 g_2 g 2 , y t y_t y t at scale k k k ), and combined multiplicities 6 and 3 counting transverse and longitudinal polarizations of W ± W^\pm W ± and Z Z Z at equal mass (no separate Debye improvement of the longitudinal modes is applied inside the flow equation itself). The top quark enters with an overall minus sign and tanh rather than coth, as required by Fermi–Dirac statistics; as τ → 0 \tau\to0 τ → 0 , tanh → 1 \tanh\to1 tanh → 1 and E e x t E_{\mathrm{ext}} E ext reduces to the usual zero-temperature top and gauge-boson threshold functions.
The dimensionless Goldstone mass is
m ˉ G 2 ( ρ , σ ) = u ρ ( ρ , σ ) . (20) \bar{m}_{G}^{2}(\rho,\sigma)=u_{\rho}(\rho,\sigma). \tag*{(20)} m ˉ G 2 ( ρ , σ ) = u ρ ( ρ , σ ) . (20)
The dimensionless radial mass matrix is given by
M ˉ 2 ( ρ , σ ) = ( u ρ + 2 ρ u ρ ρ 2 ρ σ u ρ σ 2 ρ σ u ρ σ u σ + 2 σ u σ σ ) , (21) \bar{M}^{2}(\rho,\sigma)=\begin{pmatrix} u_{\rho}+2\rho u_{\rho\rho} & 2\sqrt{\rho\sigma}\,u_{\rho\sigma} \\ 2\sqrt{\rho\sigma}\,u_{\rho\sigma} & u_{\sigma}+2\sigma u_{\sigma\sigma} \end{pmatrix}, \tag*{(21)} M ˉ 2 ( ρ , σ ) = ( u ρ + 2 ρ u ρρ 2 ρ σ u ρ σ 2 ρ σ u ρ σ u σ + 2 σ u σ σ ) , (21)
and its eigenvalues are
m ˉ ± 2 ( ρ , σ ) = 1 2 [ M ˉ 11 2 + M ˉ 22 2 ± ( M ˉ 11 2 − M ˉ 22 2 ) 2 + 4 ( M ˉ 12 2 ) 2 ] . (22) \bar{m}_{\pm}^{2}(\rho,\sigma)=\frac{1}{2}\left[\bar{M}_{11}^{2}+\bar{M}_{22}^{2}\pm\sqrt{\left(\bar{M}_{11}^{2}-\bar{M}_{22}^{2}\right)^{2}+4\left(\bar{M}_{12}^{2}\right)^{2}}\right]. \tag*{(22)} m ˉ ± 2 ( ρ , σ ) = 2 1 [ M ˉ 11 2 + M ˉ 22 2 ± ( M ˉ 11 2 − M ˉ 22 2 ) 2 + 4 ( M ˉ 12 2 ) 2 ] . (22)
In our Higgs–singlet setup, the Higgs field is directly coupled to the top quark and electroweak gauge bosons, while the singlet is not. Consequently, the corresponding wave-function renormalizations Z H , k Z_{H,k} Z H , k and Z S , k Z_{S,k} Z S , k are expected to exhibit a markedly different scale dependence. Perturbatively, one finds that the anomalous dimension η H \eta_H η H is generated already at one loop, whereas η S \eta_S η S starts only at two loops. In the renormalized variables ( ρ , σ ) (\rho,\sigma) ( ρ , σ ) the wave-function factors cancel identically from Eq. (21) : inside the last integrated momentum shell the radial propagator is Z i , k k 2 + ∂ 2 V k / ∂ ϕ i ∂ ϕ j Z_{i,k}k^2+\partial^2 V_k/\partial\phi_i\partial\phi_j Z i , k k 2 + ∂ 2 V k / ∂ ϕ i ∂ ϕ j , and rescaling by Z i , k k 2 Z_{i,k}k^2 Z i , k k 2 turns the curvature of V k V_k V k into exactly the curvature of u u u . Equation (21) is therefore already the canonically normalized radial mass matrix, with the Higgs–singlet mixing u ρ σ u_{\rho\sigma} u ρ σ fully retained through the eigenvalues m ˉ ± 2 \bar{m}_{\pm}^{2} m ˉ ± 2 ; the difference Z H , k ≠ Z S , k Z_{H,k}\ne Z_{S,k} Z H , k = Z S , k does not enter it.2 What the LPA′ ansatz does drop is the field dependence of Z H , k , Z S , k Z_{H,k},Z_{S,k} Z H , k , Z S , k and any off-diagonal kinetic term Z ρ σ ∂ μ ρ ∂ μ σ Z_{\rho\sigma}\,\partial_\mu\rho\,\partial_\mu\sigma Z ρ σ ∂ μ ρ ∂ μ σ mixing the two radial modes.3 The latter is generated by the portal coupling alongside u ρ σ u_{\rho\sigma} u ρ σ , but a constant H H H –S S S kinetic mixing is forbidden by gauge invariance, so it is field-dependent (vanishing at the origin by the Z 2 Z_2 Z 2 symmetry) and parametrically of order λ H S 2 / ( 16 π 2 ) \lambda_{HS}^{2}/(16\pi^{2}) λ H S 2 / ( 16 π 2 ) times a loop factor – the same size as η S \eta_S η S itself.
If one works in strict LPA rather than LPA′, one simply sets
η H = η S = 0. (23) \eta_H=\eta_S=0. \tag*{(23)} η H = η S = 0. (23)
For the numerical calculations, we directly evaluate the dimensionless potential u ( ρ , σ ) u(\rho,\sigma) u ( ρ , σ ) by means of physics-informed neural networks (PINNs), whereas the anomalous dimensions η H \eta_H η H and η S \eta_S η S are taken from perturbative one- and two-loop results (the leading contribution to η S \eta_S η S appears only at two-loop order).
At one loop, the Higgs anomalous dimension in the SM is given by
γ H = − 1 2 ∂ t ln Z H ≃ 1 16 π 2 ( 3 y t 2 − 9 4 g 2 2 − 3 4 g 1 2 ) , η H ≃ 2 γ H , (24) \gamma_H=-\frac{1}{2}\partial_t\ln Z_H\simeq\frac{1}{16\pi^2}\left(3y_t^2-\frac{9}{4}g_2^2-\frac{3}{4}g_1^2\right),\qquad\eta_H\simeq2\gamma_H, \tag*{(24)} γ H = − 2 1 ∂ t ln Z H ≃ 16 π 2 1 ( 3 y t 2 − 4 9 g 2 2 − 4 3 g 1 2 ) , η H ≃ 2 γ H , (24)
in the M S ‾ \overline{\mathrm{MS}} MS scheme with Landau gauge. For the singlet S S S , γ S \gamma_S γ S vanishes at one loop and is first generated at two loops through the portal and singlet quartic couplings; PyR@TE3 [23 ] gives
γ S ≃ 1 ( 16 π 2 ) 2 ( λ H S 2 + 3 λ S 2 ) , η S ≃ 2 γ S . (25) \gamma_S\simeq\frac{1}{(16\pi^2)^2}\left(\lambda_{HS}^{2}+3\lambda_S^{2}\right),\qquad\eta_S\simeq2\gamma_S. \tag*{(25)} γ S ≃ ( 16 π 2 ) 2 1 ( λ H S 2 + 3 λ S 2 ) , η S ≃ 2 γ S . (25)
Over the RG-time range used here η S ≲ 10 − 4 \eta_S\lesssim10^{-4} η S ≲ 1 0 − 4 , so its precise value (and its residual scheme dependence) has no visible effect on the results; it is retained only for completeness. Both the one-loop γ H \gamma_H γ H and the two-loop γ S \gamma_S γ S were obtained with PyR@TE3 for the model defined above; the full set of two-loop β \beta β functions is collected in Appendix § 11 .
First-order electroweak phase transition
Vacuum structure and two-step transitions
In the Z 2 \mathbb{Z}_2 Z 2 -symmetric singlet extension, the zero-temperature potential can support, in addition to the electroweak vacuum ( ρ ~ , σ ~ ) = ( v 2 / 2 , 0 ) (\tilde{\rho},\tilde{\sigma})=(v^2/2,0) ( ρ ~ , σ ~ ) = ( v 2 /2 , 0 ) , a singlet-only vacuum ( ρ ~ , σ ~ ) = ( 0 , v S 2 / 2 ) (\tilde{\rho},\tilde{\sigma})=(0,v_S^2/2) ( ρ ~ , σ ~ ) = ( 0 , v S 2 /2 ) where the Z 2 \mathbb{Z}_2 Z 2 symmetry is spontaneously broken. At high temperature both minima are typically lifted relative to the symmetric point ( ρ ~ , σ ~ ) = ( 0 , 0 ) (\tilde{\rho},\tilde{\sigma})=(0,0) ( ρ ~ , σ ~ ) = ( 0 , 0 ) , and depending on the size of λ H S \lambda_{HS} λ H S the thermal history can proceed either directly from the symmetric phase to the electroweak vacuum (one-step transition), or via an intermediate stage in which the singlet direction is populated first and the system subsequently tunnels into the electroweak vacuum (two-step transition) [8 , 10 ] .
Formally, a two-step history is characterized by a temperature range in which the singlet-broken minimum is the global minimum of the finite-T T T potential,
V ( 0 , v S ; T ) < V ( v , 0 ; T ) , (26) V(0,v_S;T)<V(v,0;T), \tag*{(26)} V ( 0 , v S ; T ) < V ( v , 0 ; T ) , (26)
followed, at lower temperature, by a swap of the global minimum,
V ( 0 , v S ; T ) > V ( v , 0 ; T ) , (27) V(0,v_S;T)>V(v,0;T), \tag*{(27)} V ( 0 , v S ; T ) > V ( v , 0 ; T ) , (27)
so that the electroweak vacuum eventually becomes the true ground state. The nature of this vacuum swap — and in particular whether the second step proceeds via a strongly first-order transition — is highly sensitive to the shape of the potential barrier separating the two minima, which in turn is controlled by the same nonperturbative infrared dynamics that the FRG flow is designed to resolve. This is in contrast to one-loop perturbative treatments, where the barrier is generated only by thermal cubic terms and is therefore parametrically small unless the portal coupling λ H S \lambda_{HS} λ H S is sizable [9 ] . Two-loop perturbative studies have shown that such corrections can shift the transition strength and critical temperature by O ( 20 % – 50 % ) \mathcal{O}(20\%\text{--}50\%) O ( 20% – 50% ) relative to the one-loop result in typical strong-transition benchmarks [11 ] , which motivates a fully nonperturbative treatment such as the one pursued here. A nonperturbative lattice study of this model, based on a perturbatively dimensionally-reduced three-dimensional effective theory, has recently mapped the strongly first-order region of its parameter space and provides the natural benchmark against which a fully nonperturbative effective potential should ultimately be tested [12 ] .
Critical and nucleation temperatures
For a given temperature T T T , the flow is integrated from the UV boundary down to the infrared scale k I R = 150 k_{\mathrm{IR}}=150 k IR = 150 GeV (§ 4.1 ), yielding the flowed potential u k I R ( ρ , σ ; T ) u_{k_{\mathrm{IR}}}(\rho,\sigma;T) u k IR ( ρ , σ ; T ) used to determine the vacuum structure. For the vacuum structure and the bounce below we work with the corresponding dimensionful potential
V k I R ( ρ ~ , σ ~ ; T ) ≡ k I R 4 u k I R ( Z H , k I R ρ ~ / k I R 2 , Z S , k I R σ ~ / k I R 2 ; T ) , (28) V_{k_{\mathrm{IR}}}(\tilde{\rho},\tilde{\sigma};T)\equiv k_{\mathrm{IR}}^4 u_{k_{\mathrm{IR}}}\left(Z_{H,k_{\mathrm{IR}}}\tilde{\rho}/k_{\mathrm{IR}}^2,\ Z_{S,k_{\mathrm{IR}}}\tilde{\sigma}/k_{\mathrm{IR}}^2;T\right), \tag*{(28)} V k IR ( ρ ~ , σ ~ ; T ) ≡ k IR 4 u k IR ( Z H , k IR ρ ~ / k IR 2 , Z S , k IR σ ~ / k IR 2 ; T ) , (28)
a function of the dimensionful invariants ( ρ ~ , σ ~ ) (\tilde{\rho},\tilde{\sigma}) ( ρ ~ , σ ~ ) in G e V 2 \mathrm{GeV}^2 GeV 2 (Sec. 2.3). The critical temperature T c T_c T c is defined as the temperature at which the two competing minima are degenerate,
V k I R ( ρ ~ E W ( T c ) , 0 ; T c ) = V k I R ( 0 , σ ~ S ( T c ) ; T c ) , (29) V_{k_{\mathrm{IR}}}(\tilde{\rho}_{\mathrm{EW}}(T_c),0;T_c)=V_{k_{\mathrm{IR}}}(0,\tilde{\sigma}_S(T_c);T_c), \tag*{(29)} V k IR ( ρ ~ EW ( T c ) , 0 ; T c ) = V k IR ( 0 , σ ~ S ( T c ) ; T c ) , (29)
where ρ ~ E W ( T ) \tilde{\rho}_{\mathrm{EW}}(T) ρ ~ EW ( T ) and σ ~ S ( T ) \tilde{\sigma}_S(T) σ ~ S ( T ) denote the location of the electroweak and singlet minima in the dimensionful invariants, tracked continuously as functions of T T T from the flowed potential. Because u k I R u_{k_{\mathrm{IR}}} u k IR is obtained directly from the PINN representation on a continuous domain, both minima and the saddle point (barrier top) separating them can be located by a simple gradient-based search on the trained network, without interpolation artifacts.
Stopping the flow at k I R = 150 k_{\mathrm{IR}}=150 k IR = 150 GeV rather than at k → 0 k\to0 k → 0 is a practical choice tied to the benchmark (Λ ≃ 1108 \Lambda\simeq1108 Λ ≃ 1108 GeV over an RG-time range t r a n g e = − 2 t_{\mathrm{range}}=-2 t range = − 2 ; § 4.1 ). Since k I R ∼ T ∼ m h k_{\mathrm{IR}}\sim T\sim m_h k IR ∼ T ∼ m h , the flow is close to frozen in the outer, deeply broken region, so the residual running below k I R k_{\mathrm{IR}} k IR is a small systematic on the vacuum locations and depths. It is not a benign one for the barrier, however: continuing to substantially lower k I R k_{\mathrm{IR}} k IR takes the inner, nearly flat region between the minima into the regime where the potential convexifies and 1 + m ˉ 2 → 0 1+\bar{m}^{2}\to0 1 + m ˉ 2 → 0 , where the coth threshold functions become singular and a controlled Maxwell-type treatment of the convexification is required – the same obstruction behind the floor on 1 + m ˉ 2 1+\bar{m}^{2} 1 + m ˉ 2 and the tachyonic abort of § 5.2 . The inner barrier is thus both the quantity most sensitive to lowering k I R k_{\mathrm{IR}} k IR and the hardest to resolve, consistent with it being the least converged output of the two-dimensional network (§ 6.3 ); propagating this systematic is left for future work.
The strength of the transition is quantified by the standard order parameter
ξ c ≡ v c T c , v c ≡ 2 ρ ~ E W ( T c ) , (30) \xi_c \equiv\frac{v_c}{T_c}, \qquad v_c \equiv\sqrt{2\widetilde{\rho}_{\mathrm{EW}}(T_c)}, \tag*{(30)} ξ c ≡ T c v c , v c ≡ 2 ρ EW ( T c ) , (30)
with ξ c ≳ 1 \xi_c \gtrsim1 ξ c ≳ 1 conventionally taken as the (parametric) condition for a transition strong enough to avoid washout of the baryon asymmetry in electroweak-baryogenesis scenarios. As in conventional effective-potential treatments, the LPA′ ^\prime ′ flow here is formulated in a fixed gauge (Landau), so ξ c \xi_c ξ c and the barrier inherit the usual gauge dependence of the effective potential; a manifestly gauge-invariant treatment is beyond the scope of this proof of concept. Because T c T_c T c only signals degeneracy of the two minima and not the actual completion of the transition, we additionally estimate the nucleation temperature T n < T c T_n<T_c T n < T c from the O ( 3 ) O(3) O ( 3 ) bounce action [24 , 25 ] ,
S 3 ( T ) = 4 π ∫ 0 ∞ d r r 2 [ 1 2 ( d ϕ d r ) 2 + Δ V k I R ( ϕ ( r ) ; T ) ] , (31) S_3(T)=4\pi\int_0^\infty\mathrm{d}r\,r^2\left[\frac{1}{2}\left(\frac{\mathrm{d}\phi}{\mathrm{d}r}\right)^2+\Delta V_{k_{\mathrm{IR}}}(\phi(r);T)\right], \tag*{(31)} S 3 ( T ) = 4 π ∫ 0 ∞ d r r 2 [ 2 1 ( d r d ϕ ) 2 + Δ V k IR ( ϕ ( r ) ; T ) ] , (31)
evaluated on the solution ϕ ( r ) = ( h ( r ) , s ( r ) ) \phi(r)=(h(r),s(r)) ϕ ( r ) = ( h ( r ) , s ( r )) of the two-field bounce equation
d 2 ϕ d r 2 + 2 r d ϕ d r = ∇ ϕ Δ V k I R ( ϕ ; T ) , ϕ ′ ( 0 ) = 0 , ϕ ( r → ∞ ) → ϕ f a l s e , (32) \frac{\mathrm{d}^2\phi}{\mathrm{d}r^2}+\frac{2}{r}\frac{\mathrm{d}\phi}{\mathrm{d}r}=\nabla_{\phi}\Delta V_{k_{\mathrm{IR}}}(\phi;T), \qquad\phi^\prime(0)=0,\qquad\phi(r\to\infty)\to\phi_{\mathrm{false}}, \tag*{(32)} d r 2 d 2 ϕ + r 2 d r d ϕ = ∇ ϕ Δ V k IR ( ϕ ; T ) , ϕ ′ ( 0 ) = 0 , ϕ ( r → ∞ ) → ϕ false , (32)
where ϕ = ( h , s ) \phi=(h,s) ϕ = ( h , s ) are the canonically normalized radial fields (ρ ~ = h 2 / 2 \widetilde{\rho}=h^2/2 ρ = h 2 /2 , σ ~ = s 2 / 2 \widetilde{\sigma}=s^2/2 σ = s 2 /2 ), ϕ f a l s e \phi_{\mathrm{false}} ϕ false is the false vacuum, and Δ V k I R \Delta V_{k_{\mathrm{IR}}} Δ V k IR is the potential measured relative to the true vacuum. We solve this for the flowed two-dimensional potential with the path-deformation method of Ref. [26 ] , which alternates between the one-dimensional bounce along a fixed field-space path and a deformation of that path toward the local force. Nucleation is taken to occur when S 3 ( T n ) / T n ≃ 140 S_3(T_n)/T_n\simeq140 S 3 ( T n ) / T n ≃ 140 (≃ 4 ln ( M P l / T ) \simeq4\ln(M_{\mathrm{Pl}}/T) ≃ 4 ln ( M Pl / T ) at the electroweak scale), the standard criterion for roughly one bubble per Hubble volume [5 , 27 ] . In this work we evaluate S 3 ( T ) / T S_3(T)/T S 3 ( T ) / T at a fixed benchmark temperature directly from the flowed two-dimensional potential (§ 6.3 ); locating T n T_n T n , and T c T_c T c , by a temperature scan would require retraining the network at each temperature and is left for future work.
Ring (Debye mass) resummation and its FRG counterpart
In perturbative treatments of the finite-temperature effective potential, the appearance of IR-sensitive daisy (ring) diagrams associated with the zero Matsubara mode requires resummation of thermal (Debye) masses in the scalar and longitudinal gauge propagators [28 , 29 ] ; without this resummation the loop expansion in the high-temperature regime breaks down order by order in the coupling. In our framework, the scalar sector of the Higgs–singlet system is treated nonperturbatively by means of the functional renormalization group, and the corresponding resummation of thermal fluctuations is generated automatically by the Matsubara sum and the successive integration of momentum shells from k = Λ k=\Lambda k = Λ down to k → 0 k\to0 k → 0 . Concretely, the bosonic contribution in the flow equation,
T ∑ n ∈ Z ⟶ k 2 π 1 + m ˉ 2 coth ( 1 + m ˉ 2 2 T ) , (33) T\sum_{n\in\mathbb{Z}}\longrightarrow\frac{k}{2\pi}\sqrt{1+\bar{m}^{2}}\coth\left(\frac{\sqrt{1+\bar{m}^{2}}}{2T}\right), \tag*{(33)} T n ∈ Z ∑ ⟶ 2 π k 1 + m ˉ 2 coth ( 2 T 1 + m ˉ 2 ) , (33)
already includes the zero-mode contribution together with all nonzero Matsubara modes, and the k k k -dependent masses m ˉ 2 \bar{m}^{2} m ˉ 2 entering this expression incorporate the effects of fluctuations integrated out at higher scales. In this sense the FRG flow for the scalar modes performs an all-orders resummation of the ring diagrams without requiring a separate insertion of thermal masses into tree-level propagators.
At the same time, the electroweak gauge bosons and the top quark are not included dynamically in the FRG flow of Γ k [ H , S ] \Gamma_k[H,S] Γ k [ H , S ] but are treated as an external perturbative sector, entering the scalar flow through the term E e x t E_{\mathrm{ext}} E ext of Eq. (19) . Inside the flow this term is evaluated with tree-level (running-coupling) gauge and top masses only, with no Debye improvement of the longitudinal modes, so its shell integration reconstructs the one-loop gauge/top contribution to the potential without the daisy resummation. The ring resummation of this external sector is instead restored a posteriori , by adding it once and explicitly to the flowed potential – the “NN+ring” construction used for the comparisons of Sec. 6 – and it is likewise built by hand into the perturbative finite-temperature reference curves. Inserting Debye masses into E e x t E_{\mathrm{ext}} E ext itself would double-count this thermal self-energy against that a posteriori correction. Ring resummation is thus unnecessary within the FRG treatment of the scalar Higgs–singlet fields, whose flow includes it automatically through the full Wetterich equation, but it remains essential for the externally supplied gauge and Yukawa contributions in order to control their high-temperature infrared behaviour consistently in our hybrid setup.
Benchmark: a grid-based Newton–Krylov relaxation solver
Before turning to the PINN construction, it is useful to record what happens if the flow equation derived above is instead attacked with a conventional grid-based method, both to motivate the mesh-free approach of § 5 and because the resulting (non-)convergence pattern is itself physically informative about the equation being solved.
Benchmark point and input parameters
The flow is followed over the RG-time range t ∈ [ t r a n g e , 0 ] t \in[t_{\mathrm{range}}, 0] t ∈ [ t range , 0 ] , with t r a n g e = − 2 t_{\mathrm{range}} = -2 t range = − 2 (the same range used for the multi-block training of § 5.3 ), corresponding to a physical scale range
k I R = 150 G e V ≤ k ≤ Λ ≡ k U V = k I R e − t r a n g e ≃ 1108 G e V . (34) k_{\mathrm{IR}} = 150\,\mathrm{GeV} \le k \le\Lambda\equiv k_{\mathrm{UV}} = k_{\mathrm{IR}} e^{-t_{\mathrm{range}}} \simeq1108\,\mathrm{GeV}. \tag*{(34)} k IR = 150 GeV ≤ k ≤ Λ ≡ k UV = k IR e − t range ≃ 1108 GeV . (34)
The Higgs-sector tree-level parameters entering Eq. (16) are not free: they are fixed by the physical Higgs pole mass and electroweak VEV, m h = 125.2 G e V m_h = 125.2\,\mathrm{GeV} m h = 125.2 GeV and v = 246.2 G e V v = 246.2\,\mathrm{GeV} v = 246.2 GeV (i.e. ⟨ h ⟩ = v \langle h \rangle= v ⟨ h ⟩ = v , ρ ˉ E W = v 2 / 2 \bar{\rho}_{\mathrm{EW}} = v^2/2 ρ ˉ EW = v 2 /2 ), through the standard tree-level relations
λ H t r e e = m h 2 2 v 2 ≃ 0.129 , μ H 2 , t r e e = − λ H t r e e v 2 = − m h 2 2 ≃ − 7838 G e V 2 . (35) \lambda_H^{\mathrm{tree}} = \frac{m_h^2}{2v^2} \simeq0.129,\qquad\mu_H^{2,\mathrm{tree}} = -\lambda_H^{\mathrm{tree}}v^2 = -\frac{m_h^2}{2} \simeq-7838\,\mathrm{GeV}^2. \tag*{(35)} λ H tree = 2 v 2 m h 2 ≃ 0.129 , μ H 2 , tree = − λ H tree v 2 = − 2 m h 2 ≃ − 7838 GeV 2 . (35)
The singlet sector, by contrast, is unconstrained by any existing measurement, and we simply choose a tree-level benchmark point
λ S t r e e = 0.27 , λ H S t r e e = 1.0 , μ S 2 , t r e e = − ( 90 G e V ) 2 = − 8100 G e V 2 , (36) \lambda_S^{\mathrm{tree}} = 0.27,\qquad\lambda_{HS}^{\mathrm{tree}} = 1.0,\qquad\mu_S^{2,\mathrm{tree}} = -(90\,\mathrm{GeV})^2 = -8100\,\mathrm{GeV}^2, \tag*{(36)} λ S tree = 0.27 , λ H S tree = 1.0 , μ S 2 , tree = − ( 90 GeV ) 2 = − 8100 GeV 2 , (36)
with λ H S \lambda_{HS} λ H S the Higgs–singlet portal coupling.
These tree-level values are then promoted to couplings renormalized at μ = k I R \mu= k_{\mathrm{IR}} μ = k IR by including the leading one-loop correction. Together with the one-loop-matched gauge couplings and a top Yukawa y t ( k I R ) ≃ 0.9 y_t(k_{\mathrm{IR}}) \simeq0.9 y t ( k IR ) ≃ 0.9 (consistent with the MS top Yukawa at this scale), so that g 1 , g 2 , g 3 , y t ≃ 0.358 , 0.649 , 1.176 , 0.9 g_1,g_2,g_3,y_t \simeq0.358, 0.649, 1.176, 0.9 g 1 , g 2 , g 3 , y t ≃ 0.358 , 0.649 , 1.176 , 0.9 at k I R k_{\mathrm{IR}} k IR , we form the one-loop Coleman–Weinberg potential built from the same tree-level bare masses as Eq. (22) (the two scalar mass eigenstates m ˉ ± 2 \bar{m}_{\pm}^2 m ˉ ± 2 , W W W , Z Z Z , and top),
u C W ( ρ , σ ) = 1 64 π 2 ∑ i n i m ˉ i 4 ( ρ , σ ) [ ln m ˉ i 2 ( ρ , σ ) − c i ] , (37) u_{\mathrm{CW}}(\rho,\sigma) = \frac{1}{64\pi^2}\sum_i n_i \bar{m}_i^4(\rho,\sigma)\left[\ln\bar{m}_i^2(\rho,\sigma)-c_i\right], \tag*{(37)} u CW ( ρ , σ ) = 64 π 2 1 i ∑ n i m ˉ i 4 ( ρ , σ ) [ ln m ˉ i 2 ( ρ , σ ) − c i ] , (37)
with dof/scheme constants ( n i , c i ) (n_i,c_i) ( n i , c i ) equal to ( 1 , 3 / 2 ) (1,3/2) ( 1 , 3/2 ) for each of m ˉ ± 2 \bar{m}_{\pm}^2 m ˉ ± 2 , ( 6 , 5 / 6 ) (6,5/6) ( 6 , 5/6 ) for W W W , ( 3 , 5 / 6 ) (3,5/6) ( 3 , 5/6 ) for Z Z Z , and ( − 12 , 3 / 2 ) (-12,3/2) ( − 12 , 3/2 ) for the top (the minus sign reflecting Fermi–Dirac statistics), all evaluated at μ = k = k I R \mu= k = k_{\mathrm{IR}} μ = k = k IR . The Goldstone loop is deliberately left out of this sum: at the tree-level electroweak point the Goldstone mass m ˉ G 2 \bar{m}_G^2 m ˉ G 2 vanishes exactly [Eq. (20) evaluated at ( ρ , σ ) = ( v 2 / 2 k I R 2 , 0 ) (\rho,\sigma) = (v^2/2k_{\mathrm{IR}}^2,0) ( ρ , σ ) = ( v 2 /2 k IR 2 , 0 ) gives m ˉ G 2 = 0 \bar{m}_G^2 = 0 m ˉ G 2 = 0 by construction, since this is precisely the direction of the spontaneously broken generator], so its contribution to the curvature of u C W u_{\mathrm{CW}} u CW (though not its value) hits the well-known Goldstone-boson-catastrophe log divergence [30 , 31 ] :
∂ 2 ∂ m G 2 [ m G 2 ( ln m G 2 − c ) ] ∝ ln m G 2 → − ∞ \frac{\partial^2}{\partial m_G^2}\left[m_G^2(\ln m_G^2-c)\right] \propto\ln m_G^2 \to-\infty ∂ m G 2 ∂ 2 [ m G 2 ( ln m G 2 − c ) ] ∝ ln m G 2 → − ∞
as m G 2 → 0 m_G^2 \to0 m G 2 → 0 . In a numerical implementation this divergence is only cut off by whatever ad hoc infrared regulator is used to keep ln ∣ m G 2 ∣ \ln|m_G^2| ln ∣ m G 2 ∣ finite, and the resulting matched couplings are highly sensitive to that regulator rather than converged; the other four loops above are all evaluated at genuinely nonzero tree masses and are unaffected. We therefore only include the Goldstone contribution, as usual, in the full finite-temperature potential used elsewhere in this paper (§ 3 , and the full flow equation solved in the remainder of this section), where its thermal Debye mass regulates it, and not in this T = 0 T = 0 T = 0 input-parameter matching step. Evaluating the first and second derivatives of u C W u_{\mathrm{CW}} u CW at the electroweak point ( ρ , σ ) = ( v 2 / ( 2 k I R 2 ) , 0 ) (\rho,\sigma) = (v^2/(2k_{\mathrm{IR}}^2),0) ( ρ , σ ) = ( v 2 / ( 2 k IR 2 ) , 0 ) and subtracting them from the tree-level parameters defines the renormalized couplings at k I R k_{\mathrm{IR}} k IR ,
λ H ( k I R ) = λ H t r e e − 1 2 ∂ ρ 2 u C W ∣ E W , μ H 2 ( k I R ) = μ H 2 , t r e e − k I R 2 ∂ ρ u C W ∣ E W , λ S ( k I R ) = λ S t r e e − 1 2 ∂ σ 2 u C W ∣ E W , μ S 2 ( k I R ) = μ S 2 , t r e e − k I R 2 ∂ σ u C W ∣ E W , λ H S ( k I R ) = λ H S t r e e − ∂ ρ ∂ σ u C W ∣ E W , (38) \begin{aligned}
\lambda_H(k_{\mathrm{IR}}) &= \lambda_H^{\mathrm{tree}}-\left.\frac{1}{2}\partial_\rho^2u_{\mathrm{CW}}\right|_{\mathrm{EW}}, & \mu_H^2(k_{\mathrm{IR}}) &= \mu_H^{2,\mathrm{tree}}-k_{\mathrm{IR}}^2\left.\partial_\rho u_{\mathrm{CW}}\right|_{\mathrm{EW}},\\
\lambda_S(k_{\mathrm{IR}}) &= \lambda_S^{\mathrm{tree}}-\left.\frac{1}{2}\partial_\sigma^2u_{\mathrm{CW}}\right|_{\mathrm{EW}}, & \mu_S^2(k_{\mathrm{IR}}) &= \mu_S^{2,\mathrm{tree}}-k_{\mathrm{IR}}^2\left.\partial_\sigma u_{\mathrm{CW}}\right|_{\mathrm{EW}},\\
\lambda_{HS}(k_{\mathrm{IR}}) &= \lambda_{HS}^{\mathrm{tree}}-\left.\partial_\rho\partial_\sigma u_{\mathrm{CW}}\right|_{\mathrm{EW}}, \tag*{(38)}
\end{aligned} λ H ( k IR ) λ S ( k IR ) λ H S ( k IR ) = λ H tree − 2 1 ∂ ρ 2 u CW EW , = λ S tree − 2 1 ∂ σ 2 u CW EW , = λ H S tree − ∂ ρ ∂ σ u CW ∣ EW , μ H 2 ( k IR ) μ S 2 ( k IR ) = μ H 2 , tree − k IR 2 ∂ ρ u CW ∣ EW , = μ S 2 , tree − k IR 2 ∂ σ u CW ∣ EW , (38)
i.e. each renormalized parameter is the tree value minus the corresponding curvature of the one-loop effective potential, so that once the one-loop piece is added back the renormalized point reproduces the physical Higgs mass and the chosen singlet benchmark, rather than only at tree level. The k I R k_{\mathrm{IR}} k IR column of Table 1 lists the resulting λ H ( k I R ) \lambda_H(k_{\mathrm{IR}}) λ H ( k IR ) , λ S ( k I R ) \lambda_S(k_{\mathrm{IR}}) λ S ( k IR ) , λ H S ( k I R ) \lambda_{HS}(k_{\mathrm{IR}}) λ H S ( k IR ) , μ H 2 ( k I R ) \mu_H^2(k_{\mathrm{IR}}) μ H 2 ( k IR ) , μ S 2 ( k I R ) \mu_S^2(k_{\mathrm{IR}}) μ S 2 ( k IR ) . Starting from this point at μ 0 = k I R \mu_0 = k_{\mathrm{IR}} μ 0 = k IR , the same five couplings, together with the gauge and top Yukawa couplings, are evolved with the two-loop RGEs (Appendix § 11 ) up to μ = Λ = k U V \mu= \Lambda= k_{\mathrm{UV}} μ = Λ = k UV ; the resulting values fix a H a_H a H , a S a_S a S , λ H \lambda_H λ H , λ S \lambda_S λ S , λ H S \lambda_{HS} λ H S in the UV tree potential of Eq. (16) and make up the second column of Table 1 . The same running couplings, evaluated at intermediate t t t , also supply the t t t -dependent tree-level reference entering the analytic seed potential of Eq. (50) and the perturbative comparison curves used throughout the paper.
Table 1. Higgs–singlet quartic couplings and mass parameters at the IR end of the flow (after the one-loop Coleman–Weinberg matching described in the text) and at the UV end (after two-loop RGE running), used to fix a H a_H a H , a S a_S a S , λ H \lambda_H λ H , λ S \lambda_S λ S , λ H S \lambda_{HS} λ H S in the UV boundary potential of Eq. (16) . At Λ \Lambda Λ these correspond to the dimensionless a H = μ H 2 / Λ 2 ≃ − 0.0072 a_H = \mu_H^2/\Lambda^2 \simeq-0.0072 a H = μ H 2 / Λ 2 ≃ − 0.0072 and a S = μ S 2 / Λ 2 ≃ − 0.0070 a_S = \mu_S^2/\Lambda^2 \simeq-0.0070 a S = μ S 2 / Λ 2 ≃ − 0.0070 .
Method
Rather than integrating the flow forward in t t t from the UV boundary condition by explicit or adaptive shooting, we discretize scale and field space together and solve for the entire trajectory U ( t , ρ , σ ) U(t,\rho,\sigma) U ( t , ρ , σ ) as a single global system: field derivatives are taken with a second-order central-difference stencil on a uniform ( ρ , σ ) (\rho,\sigma) ( ρ , σ ) grid, so that R H S ( t , U ) \mathrm{RHS}(t,U) RHS ( t , U ) below denotes the right-hand side of the flow equation (17) – already written with ∂ t u \partial_t u ∂ t u isolated on the left – evaluated as a function of U U U and its central-difference field derivatives at fixed t t t on the grid. The t t t direction is then discretized with Crank–Nicolson, so that the residual
R n ( U 1 , … , U N t ) = ( U n − U n − 1 ) − Δ t 2 [ RHS ( t n − 1 , U n − 1 ) + RHS ( t n , U n ) ] = 0 , n = 1 , … , N t , (39) R^n(U^1,\ldots,U^{N_t}) = \left(U^n-U^{n-1}\right)-\frac{\Delta t}{2}\left[\operatorname{RHS}(t^{n-1},U^{n-1})+\operatorname{RHS}(t^n,U^n)\right]
=0,\qquad n=1,\ldots,N_t, \tag*{(39)} R n ( U 1 , … , U N t ) = ( U n − U n − 1 ) − 2 Δ t [ RHS ( t n − 1 , U n − 1 ) + RHS ( t n , U n ) ] = 0 , n = 1 , … , N t , (39)
with U 0 U^0 U 0 fixed to the UV tree-level potential u Λ u_\Lambda u Λ of Eq. (16) , is required to vanish simultaneously at every time slice. In the tree-only limit (η = 0 \eta=0 η = 0 , no E e x t E_{\mathrm{ext}} E ext /coth term) this system is linear in U U U , so Newton iteration is unnecessary and the block bidiagonal system is solved exactly by a single LU factorization; validated against the exact characteristic-curve solution, this scheme reduces the relative error from 27%, obtained with the original upwind/shooting discretization, to 0.02%–0.6% depending on grid resolution, showing that the dominant source of error in the shooting approach was the first-order upwind spatial stencil rather than the marching direction itself.
Once the loop term E e x t E_{\mathrm{ext}} E ext and the coth threshold functions are switched on the system becomes nonlinear, and we solve it with a Jacobian-free Newton–Krylov method [32 ] (the newton_krylov routine from scipy.optimize, with an LGMRES [33 ] inner solver), seeded by the closed-form sum of the tree solution and a fixed coarse Debye ansatz for the thermal piece. This is the regime in which a relaxation method could in principle realize its main advantage over shooting: convergence of the entire trajectory to a single self-consistent fixed point, rather than the one-directional accumulation of error characteristic of forward marching.
A first check: the Higgs-only reduction (relaxation solver)
Before attacking the full two-field problem it is worth checking the solver on the simplest version of it: the singlet-free reduction of the model – Higgs plus the same external gauge and top sector – in which the radial sector is one-dimensional. The only scalar masses are then the Goldstone mass m ˉ G 2 = u ρ \bar{m}_G^2=u_\rho m ˉ G 2 = u ρ and the single radial mass m ˉ H 2 = u ρ + 2 ρ u ρ ρ \bar{m}_H^2=u_\rho+2\rho u_{\rho\rho} m ˉ H 2 = u ρ + 2 ρ u ρρ , the loop sum reducing to 3 / 1 + m ˉ G 2 coth ( ⋯ ) + 1 / 1 + m ˉ H 2 coth ( ⋯ ) 3/\sqrt{1+\bar{m}_G^2}\,\coth(\cdots)+1/\sqrt{1+\bar{m}_H^2}\,\coth(\cdots) 3/ 1 + m ˉ G 2 coth ( ⋯ ) + 1/ 1 + m ˉ H 2 coth ( ⋯ ) , so that the 2 × 2 2\times2 2 × 2 radial mass matrix of Eq. (21) , its near-degenerate eigenvalue mixing and the associated disc \sqrt{\operatorname{disc}} disc – a natural first suspect for the stall of the full two-field Newton–Krylov iteration (§ 4.4 ) – are absent by construction. The regulator, the one-loop Coleman–Weinberg matching, the two-loop running couplings, the finite-temperature treatment, the central-difference/Crank–Nicolson discretization and the Newton–Krylov solver are all kept identical.
Removing the singlet does not, however, make the loop-level flow easy to solve reliably, and the behaviour differs sharply between the two benchmark temperatures. At T = 100 T=100 T = 100 GeV the unfoced baseline Newton–Krylov iteration does reach the requested tolerance f t o l = 10 − 8 f_{\mathrm{tol}}=10^{-8} f tol = 1 0 − 8 (“converged” by the solver’s own criterion), but the solution it converges to is unphysical: it decreases monotonically to u ≃ − 0.96 u\simeq-0.96 u ≃ − 0.96 at ρ = ρ max = 1.75 \rho=\rho_{\max}=1.75 ρ = ρ m a x = 1.75 , far below both the tree-level potential (u = − 0.37 u=-0.37 u = − 0.37 there) and the ring-resummed (Arnold–Espinosa) finite-temperature reference (u = − 0.06 u=-0.06 u = − 0.06 ) – a spurious deep well rather than the physical potential. A small residual therefore does not certify a physically correct solution even in this single-field case. At T = 200 T=200 T = 200 GeV the iteration instead does not converge at all: the residual norm ∥ R ∥ 2 \lVert R\rVert_2 ∥ R ∥ 2 stalls at 6.8 × 10 − 3 6.8\times10^{-3} 6.8 × 1 0 − 3 (below the 0.14 0.14 0.14 –0.73 0.73 0.73 range found for the full two-field system below, Table 2 , though part of that gap simply reflects the smaller one-dimensional grid), and the stalled solution undershoots the finite-temperature reference badly (u = − 0.43 u=-0.43 u = − 0.43 at ρ max \rho_{\max} ρ m a x versus u = + 0.55 u=+0.55 u = + 0.55 ), missing essentially the entire thermal lift of the potential (Figure 1 ). Adding the coupling-prior forcing term defined in § 4.5 below – a soft perturbative-consistency penalty added directly to the residual, here supplemented by an explicit finite-temperature target on the mass-like combination – pulls the T = 200 T=200 T = 200 solution toward the physical reference, raising u ( ρ max ) u(\rho_{\max}) u ( ρ m a x ) to + 0.24 +0.24 + 0.24 , but it still undershoots by roughly a factor of two and raises the residual norm to 4.0 × 10 − 2 4.0\times10^{-2} 4.0 × 1 0 − 2 .
Figure 1. First check on the Higgs-only reduction of the grid-based benchmark at T = 200 T = 200 T = 200 GeV and the IR end of the flow t = t r a n g e = − 2 t = t_{\mathrm{range}} = -2 t = t range = − 2 . The unforced Newton–Krylov solution of the full flow equation (orange squares) undershoots the ring-resummed (Arnold–Espinosa) finite-temperature reference (green dotted, the closest available analogue of the true nonperturbative trajectory) badly, missing essentially the entire thermal lift; the tree-level and zero-temperature Coleman–Weinberg curves (black dashed, grey dash-dotted) are shown for context. Adding the coupling-prior forcing term with an explicit finite-temperature mass target (purple triangles) moves the solution toward that reference but still undershoots it by about a factor of two, while raising the residual norm ∥ R ∥ 2 \lVert R\rVert_2 ∥ R ∥ 2 from 6.8 × 10 − 3 6.8 \times10^{-3} 6.8 × 1 0 − 3 to 4.0 × 10 − 2 4.0 \times10^{-2} 4.0 × 1 0 − 2 . All curves are shifted so that u = 0 u = 0 u = 0 at ρ = 0 \rho= 0 ρ = 0 . (At T = 100 T = 100 T = 100 GeV, not shown, the unforced solver instead reaches tolerance but converges to a spurious deep well; see the text.)
Table 2. Final residual norm ∥ R ∥ 2 \lVert R\rVert_2 ∥ R ∥ 2 and wall-clock time of the Jacobian-free Newton–Krylov relaxation on a 21 × 21 21 \times21 21 × 21 spatial grid with N t = 40 N_t = 40 N t = 40 Crank–Nicolson steps, for the four variants shown in Figure 2 –Figure 3 . All four use the domain restriction ρ , σ ≥ 10 − 3 \rho,\sigma\ge10^{-3} ρ , σ ≥ 1 0 − 3 of § 4.4 with disc_cut = 0; the coupling-prior variants additionally add the forcing term of § 4.5 with weights ( 0.015 , 0.006 ) (0.015, 0.006) ( 0.015 , 0.006 ) . All runs use f t o l = 10 − 8 f_{\mathrm{tol}} = 10^{-8} f tol = 1 0 − 8 and a maximum of 100 outer Newton iterations. None of the eight runs reaches the requested tolerance.
Figure 2. Slices of the effective potential u ( t r a n g e , ρ , σ ) u(t_{\mathrm{range}},\rho,\sigma) u ( t range , ρ , σ ) along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 (left) and σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 (right) at T = 100 G e V T = 100\ \mathrm{GeV} T = 100 GeV , with the domain restricted to ρ , σ ≥ 10 − 3 \rho,\sigma\ge10^{-3} ρ , σ ≥ 1 0 − 3 , comparing the finite-difference relaxation solution of the tree-only flow (numerical, blue) against its exact counterpart (tree, black dashed), a static one-loop Coleman–Weinberg reference (grey dash-dotted), the same reference with the Arnold–Espinosa ring-resummed finite-temperature one-loop potential added (green dotted, “physical reference”), and the four non-converged Newton–Krylov variants of the full flow equation described in the text (orange, purple, brown, pink). All curves are shifted so that u = 0 u = 0 u = 0 at ( ρ , σ ) = ( 10 − 3 , 10 − 3 ) (\rho,\sigma) = (10^{-3},10^{-3}) ( ρ , σ ) = ( 1 0 − 3 , 1 0 − 3 ) .
Figure 3. Same as Figure 2 , at T = 200 T = 200 T = 200 GeV. Along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 (left) the coupling-prior and fourth-order-plus-prior variants (purple, pink) lie almost on top of each other, with the fourth-order variant without the prior (brown) sitting between them and the unforced baseline (orange); along σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 (right) this separation largely disappears and all four variants, including the baseline, cluster within ∼ 20 % \sim20\% ∼ 20% of each other while undershooting the thermally lifted reference curve (green dotted) by roughly a factor of two.
So neither temperature yields a trustworthy single-field solution: a converged-but-spurious one at T = 100 T = 100 T = 100 GeV, an unconverged one at T = 200 T = 200 T = 200 GeV. Since the reduction contains none of the singlet mass-matrix structure, this difficulty cannot be attributed to the eigenvalue mixing – it is intrinsic to the nonlinear loop-plus-thermal flow equation. The full two-field system, to which we now turn, adds the near-degenerate eigenvalue mixing of Eq. (22) along the symmetry axes; we find below that this near-degeneracy is not what obstructs the solver either. The same Higgs-only reduction is revisited for the PINN in § 6.1 , with a different outcome.
Convergence failure and its origin
Turning to the full two-field system, Newton–Krylov iteration on the full equation does not converge to the requested tolerance (f t o l = 10 − 8 f_{\mathrm{tol}} = 10^{-8} f tol = 1 0 − 8 )4 for any of the regularization choices we tried, on a 21 × 21 × 40 21 \times21 \times40 21 × 21 × 40 grid: the residual norm drops rapidly during the first few iterations (by ≳ 95 % \gtrsim95\% ≳ 95% relative to the initial warm start) and then stalls at a plateau of order 10 − 1 10^{-1} 1 0 − 1 , with the point-wise maximum residual settling around 6 × 10 − 3 6 \times10^{-3} 6 × 1 0 − 3 –1 × 10 − 1 1 \times10^{-1} 1 × 1 0 − 1 depending on scheme (Table 2 ). Backward-Euler time discretization is markedly worse than Crank–Nicolson, with the residual norm occasionally increasing over the course of the iteration rather than decreasing monotonically.
The mechanism one would suspect first – the d i s c \sqrt{\mathrm{disc}} disc in the radial mass eigenvalues m ˉ ± 2 \bar{m}^{2}_{\pm} m ˉ ± 2 of Eq. (22) , non-analytic where the two radial eigenstates become degenerate (d i s c → 0 \mathrm{disc} \to0 disc → 0 ) – is not the cause. The eigenvalue pair enters the loop only through the symmetric combination h ( m ˉ + 2 ) + h ( m ˉ − 2 ) h(\bar{m}^{2}_{+}) + h(\bar{m}^{2}_{-}) h ( m ˉ + 2 ) + h ( m ˉ − 2 ) , with h ( x ) ≡ ( 1 + x ) − 1 / 2 coth ( 1 + x / 2 τ ) h(x) \equiv(1+x)^{-1/2}\coth\left(\sqrt{1+x}/2\tau\right) h ( x ) ≡ ( 1 + x ) − 1/2 coth ( 1 + x /2 τ ) , in which d i s c \sqrt{\mathrm{disc}} disc cancels between the two terms; the only genuine non-analyticity of h h h is at 1 + x = 0 1+x=0 1 + x = 0 , which stopping the flow at k I R = 150 G e V k_{\mathrm{IR}}=150\ \mathrm{GeV} k IR = 150 GeV keeps well outside the domain ( 1 + m ˉ 2 ≳ 0.6 (1+\bar{m}^{2} \gtrsim0.6 ( 1 + m ˉ 2 ≳ 0.6 on the tree warm start). Consistently, neither a floor on 1 + m ˉ 2 1+\bar{m}^{2} 1 + m ˉ 2 (mass_floor) nor a hard clip of d i s c \sqrt{\mathrm{disc}} disc (disc_cut) improves convergence; the runs reported here take disc_cut = 0 =0 = 0 , and under the domain restriction below disc stays above ∼ 2 × 10 − 5 \sim2 \times10^{-5} ∼ 2 × 1 0 − 5 everywhere in any case.
What limits the solver is instead poor conditioning near the symmetry axes together with the nonlinearity of the flow: Eq. (17) is first order in t t t – an ordinary initial-value problem in that variable – but fully nonlinear in the field derivatives that build the mass matrix M ˉ 2 \bar{M}^{2} M ˉ 2 of Eq. (21) : u ρ ρ u_{\rho\rho} u ρρ , u σ σ u_{\sigma\sigma} u σ σ , u ρ σ u_{\rho\sigma} u ρ σ (together with u ρ u_{\rho} u ρ , u σ u_{\sigma} u σ ) enter M ˉ 2 \bar{M}^{2} M ˉ 2 linearly, but M ˉ 2 \bar{M}^{2} M ˉ 2 ’s eigenvalues m ˉ ± 2 \bar{m}^{2}_{\pm} m ˉ ± 2 then enter the square root and coth threshold functions nonlinearly.5 The stalled residual is largest near ρ = 0 \rho=0 ρ = 0 and σ = 0 \sigma=0 σ = 0 , where the invariant chart ( ρ , σ ) = ( h ~ 2 / 2 , s ~ 2 / 2 ) (\rho,\sigma)=(\tilde{h}^{2}/2,\tilde{s}^{2}/2) ( ρ , σ ) = ( h ~ 2 /2 , s ~ 2 /2 ) is singular and the one-sided second-derivative stencils and the ρ → 0 \rho\to0 ρ → 0 regularity condition are stiffest; restricting the domain to ρ , σ ≥ 10 − 3 \rho,\sigma\ge10^{-3} ρ , σ ≥ 1 0 − 3 is the single most effective mitigation we tried, appreciably lowering the final residual norm. Reformulating the system in first-order form – auxiliary fields R ≡ u ρ R \equiv u_{\rho} R ≡ u ρ , S ≡ u σ S \equiv u_{\sigma} S ≡ u σ , so that no second-derivative stencil is evaluated anywhere – did not help further: the final residual norm was no better than the direct formulation, with or without the restriction. No combination of these measures reaches f t o l f_{\mathrm{tol}} f tol . As discussed in § 4.6 , this is because the nonlinear residual has spurious near-stationary points close to the true trajectory, so a small pointwise residual does not by itself certify the physically correct solution.
This is the mitigation the comparison of § 4.6 below actually uses: all four variants of Table 2 are run with the domain restricted to ρ , σ ≥ 10 − 3 \rho,\sigma\ge10^{-3} ρ , σ ≥ 1 0 − 3 and disc_cut = 0 =0 = 0 , and the slices shown there are taken along the resulting inner edge of the domain, ρ = 10 − 3 \rho=10^{-3} ρ = 1 0 − 3 and σ = 10 − 3 \sigma=10^{-3} σ = 1 0 − 3 , rather than along the excluded axes ρ = 0 , σ = 0 \rho=0,\sigma=0 ρ = 0 , σ = 0 . The coupling-prior forcing term introduced next is then tested as a second, independent mitigation on top of this common domain restriction, rather than as an alternative to it.
Coupling-prior forcing term
Two of the four variants compared below (§ 4.6 ) add a coupling-prior forcing term directly to the relaxation residual R n R^n R n at every grid point and time slice, implementing a soft perturbative-consistency constraint (an analogous construction is used as a PINN loss in § 5.2 , there built from the network’s derivative heads instead of finite differences).
At each grid point, local effective quartic couplings and mass parameters are read off from the current iterate U U U using the same finite-difference stencil already used for the flow operator,
λ H g r i d ≡ 1 2 u ρ ρ , λ S g r i d ≡ 1 2 u σ σ , λ H S g r i d ≡ u ρ σ , (40) \lambda_{H}^{\mathrm{grid}} \equiv\frac{1}{2}u_{\rho\rho}, \qquad\lambda_{S}^{\mathrm{grid}} \equiv\frac{1}{2}u_{\sigma\sigma}, \qquad\lambda_{HS}^{\mathrm{grid}} \equiv u_{\rho\sigma}, \tag*{(40)} λ H grid ≡ 2 1 u ρρ , λ S grid ≡ 2 1 u σ σ , λ H S grid ≡ u ρ σ , (40)
a H g r i d ≡ u ρ − u ρ ρ ρ − u ρ σ σ , a S g r i d ≡ u σ − u σ σ σ − u ρ σ ρ . (41) a_{H}^{\mathrm{grid}} \equiv u_{\rho} - u_{\rho\rho}\rho- u_{\rho\sigma}\sigma, \qquad a_{S}^{\mathrm{grid}} \equiv u_{\sigma} - u_{\sigma\sigma}\sigma- u_{\rho\sigma}\rho. \tag*{(41)} a H grid ≡ u ρ − u ρρ ρ − u ρ σ σ , a S grid ≡ u σ − u σ σ σ − u ρ σ ρ . (41)
For each of these five combinations X ∈ { λ H , λ S , λ H S , a H , a S } X \in\{\lambda_H,\lambda_S,\lambda_{HS},a_H,a_S\} X ∈ { λ H , λ S , λ H S , a H , a S } , the grid-extracted shift away from its UV tree-level value, Δ X g r i d ≡ X g r i d − X t r e e \Delta X_{\mathrm{grid}} \equiv X_{\mathrm{grid}}-X_{\mathrm{tree}} Δ X grid ≡ X grid − X tree , is compared against the corresponding perturbative shift Δ X p e r t ( t ) \Delta X_{\mathrm{pert}}(t) Δ X pert ( t ) obtained from the two-loop running couplings (Appendix § 11 ), through two hinge penalties,
ℓ s i g n [ X ] = s β ( − sign ( Δ X p e r t ) Δ X g r i d ∣ X t r e e ∣ ) , (42) \ell_{\mathrm{sign}}[X] = s_\beta\left(-\operatorname{sign}(\Delta X_{\mathrm{pert}})\frac{\Delta X_{\mathrm{grid}}}{\lvert X_{\mathrm{tree}}\rvert}\right), \tag*{(42)} ℓ sign [ X ] = s β ( − sign ( Δ X pert ) ∣ X tree ∣ Δ X grid ) , (42)
ℓ m a g [ X ] = s β ( c − ∣ Δ X p e r t ∣ − ∣ Δ X g r i d ∣ ∣ X t r e e ∣ ) + s β ( ∣ Δ X g r i d ∣ − c + ∣ Δ X p e r t ∣ ∣ X t r e e ∣ ) , (43) \ell_{\mathrm{mag}}[X] = s_\beta\left(\frac{c_-\lvert\Delta X_{\mathrm{pert}}\rvert-\lvert\Delta X_{\mathrm{grid}}\rvert}{\lvert X_{\mathrm{tree}}\rvert}\right)+s_\beta\left(\frac{\lvert\Delta X_{\mathrm{grid}}\rvert-c_+\lvert\Delta X_{\mathrm{pert}}\rvert}{\lvert X_{\mathrm{tree}}\rvert}\right), \tag*{(43)} ℓ mag [ X ] = s β ( ∣ X tree ∣ c − ∣ Δ X pert ∣ − ∣ Δ X grid ∣ ) + s β ( ∣ X tree ∣ ∣ Δ X grid ∣ − c + ∣ Δ X pert ∣ ) , (43)
with ( c − , c + ) = ( 0.1 , 3.0 ) (c_-,c_+)=(0.1,3.0) ( c − , c + ) = ( 0.1 , 3.0 ) , where s β ( x ) ≡ β − 1 ln ( 1 + e β x ) s_\beta(x)\equiv\beta^{-1}\ln(1+e^{\beta x}) s β ( x ) ≡ β − 1 ln ( 1 + e β x ) is a softplus hinge – a smooth C ∞ C^\infty C ∞ stand-in for ReLU, recovered as β → ∞ \beta\to\infty β → ∞ . Here a smooth hinge is not optional: the kink of a hard ReLU at the constraint boundary acts, for the Jacobian-free Newton–Krylov solve, as exactly the kind of non-smooth forcing that stalls the iteration (§ 4.4 ), and with a fixed nonzero weight the prior fails to converge at all unless the corner is removed. We use β = 10 \beta=10 β = 10 , which keeps the transition width within the O ( 0.1 ) \mathcal{O}(0.1) O ( 0.1 ) natural scale of the normalized arguments. ℓ s i g n [ X ] \ell_{\mathrm{sign}}[X] ℓ sign [ X ] penalizes a grid-extracted shift with the wrong sign relative to the perturbative one, while ℓ m a g [ X ] \ell_{\mathrm{mag}}[X] ℓ mag [ X ] penalizes a magnitude more than a factor 10 below or a factor 3 above the perturbative shift. Both are switched off by a relative-size mask wherever ∣ Δ X p e r t ∣ \lvert\Delta X_{\mathrm{pert}}\rvert ∣ Δ X pert ∣ is itself negligible compared to X t r e e X_{\mathrm{tree}} X tree , and by a smooth sigmoid mask in ( ρ , σ ) (\rho,\sigma) ( ρ , σ ) that confines the penalty to a neighborhood of the field-space origin where perturbation theory is expected to be reliable (in practice essentially the full 21 × 21 21\times21 21 × 21 grid used here). At the finite temperatures considered below, the sign hinge ℓ s i g n \ell_{\mathrm{sign}} ℓ sign is evaluated for all five combinations, while the magnitude band ℓ m a g \ell_{\mathrm{mag}} ℓ mag is applied only to the two mass-like combinations a H , a S a_H,a_S a H , a S ; the three quartic combinations are instead only bounded not to exceed the tree-level value by more than a factor quartic_ratio_cap = 0.99 =0.99 = 0.99 , since a clean perturbative target for the curvature of u u u is not available once nonanalytic thermal contributions are present.
The weighted sum of these penalties over the five combinations,
w s i g n ∑ X ℓ s i g n [ X ] + w m a g ∑ X ℓ m a g [ X ] , (44) w_{\mathrm{sign}}\sum_X\ell_{\mathrm{sign}}[X]+w_{\mathrm{mag}}\sum_X\ell_{\mathrm{mag}}[X], \tag*{(44)} w sign X ∑ ℓ sign [ X ] + w mag X ∑ ℓ mag [ X ] , (44)
is evaluated pointwise – independently at every grid point and time slice, rather than averaged over sampled collocation points – and the resulting value itself is added directly to RHS ( t n , U n ) \operatorname{RHS}(t^n,U^n) RHS ( t n , U n ) at that point, occupying the same slot as E e x t E_{\mathrm{ext}} E ext in Eq. (17) (rather than being back-propagated through a network as a loss gradient, since here there are no network weights to update – U U U itself is the solution). The forcing term is added with small weights ( w s i g n , w m a g ) = ( 0.015 , 0.006 ) (w_{\mathrm{sign}},w_{\mathrm{mag}})=(0.015,0.006) ( w sign , w mag ) = ( 0.015 , 0.006 ) : applied pointwise, it competes directly with the PDE residual at every single grid point rather than being averaged over a training batch as in the PINN loss, so a much smaller weight suffices. In short, this lets us test, within a purely grid-based Newton–Krylov solve, whether the same mechanism that keeps the PINN solution close in sign and order of magnitude to the perturbative expectation near the field-space origin can also stabilize (or at least regularize) the stalled relaxation iteration described above.
Comparison at T = 100 T=100 T = 100 and T = 200 T=200 T = 200 GeV
Figures 2 and 3 compare, at the IR end of the flow (t = t r a n g e = − 2 t=t_{\mathrm{range}}=-2 t = t range = − 2 ) and along ρ = 10 − 3 \rho=10^{-3} ρ = 1 0 − 3 (the singlet direction, with the Higgs field held near zero) and σ = 10 − 3 \sigma=10^{-3} σ = 1 0 − 3 (the Higgs direction), the two inner edges of the restricted domain, for variants of the (non-converged) Newton–Krylov solution of the full flow equation, all run with the domain restriction ρ , σ ≥ 10 − 3 \rho,\sigma\ge10^{-3} ρ , σ ≥ 1 0 − 3 and disc_cut = 0 =0 = 0 of § 4.4 – a baseline run using the second-order central stencil with no forcing term; the same second-order stencil with the coupling-prior forcing term of § 4.5 added, with weights ( 0.015 , 0.006 ) (0.015,0.006) ( 0.015 , 0.006 ) ; a fourth-order central spatial stencil with no forcing term; and the fourth-order stencil combined with the forcing term – against three reference curves that do not rely on solving the full equation: the exact tree-level solution, the tree solution supplemented by a static one-loop Coleman–Weinberg correction evaluated on the running couplings, and the same supplemented additionally by the Arnold–Espinosa ring-resummed finite-temperature one-loop potential (the same physical reference used for the PINN in § 6 ). Final residual norms and wall-clock times for all eight runs (two temperatures times four variants) are collected in Table 2 .
The comparison illustrates three points. First, the unforced second-order baseline is a clear outlier at T = 100 T = 100 T = 100 GeV: it reaches u = 1.15 u = 1.15 u = 1.15 at ( ρ , σ ) = ( 10 − 3 , σ max ) (\rho,\sigma) = (10^{-3},\sigma_{\max}) ( ρ , σ ) = ( 1 0 − 3 , σ m a x ) , more than four times the reference value u = 0.25 u = 0.25 u = 0.25 there, while along σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 it lands at u = − 0.54 u = -0.54 u = − 0.54 against a reference of u ≃ − 0.05 u \simeq-0.05 u ≃ − 0.05 . At T = 200 T = 200 T = 200 GeV the baseline is a less extreme outlier: along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 it overshoots the reference only mildly (u = 0.71 u = 0.71 u = 0.71 versus u = 0.46 u = 0.46 u = 0.46 , a factor ∼ 1.5 \sim1.5 ∼ 1.5 ), while along σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 it undershoots by a factor of ∼ 2.8 \sim2.8 ∼ 2.8 (u = 0.22 u = 0.22 u = 0.22 versus a reference of u = 0.60 u = 0.60 u = 0.60 ) – the same direction of failure as the other three variants there (see below), rather than the qualitatively different behaviour at T = 100 T = 100 T = 100 GeV, where the unforced baseline is a gross outlier from both the reference and the tightly clustered other three variants (by a factor of three or more). Second, the fate of the other three variants is temperature- and slice-dependent. At T = 100 T = 100 T = 100 GeV, all three – second-order-with-prior, fourth-order-without-prior, and fourth-order-with-prior – collapse onto essentially one curve regardless of whether the coupling prior is present, agreeing to within a few percent of each other pointwise (u ( 10 − 3 , σ max ) = 0.370 , 0.371 , 0.371 u(10^{-3},\sigma_{\max}) = 0.370, 0.371, 0.371 u ( 1 0 − 3 , σ m a x ) = 0.370 , 0.371 , 0.371 and u ( ρ max , 10 − 3 ) = − 0.198 , − 0.193 , − 0.196 u(\rho_{\max},10^{-3}) = -0.198, -0.193, -0.196 u ( ρ m a x , 1 0 − 3 ) = − 0.198 , − 0.193 , − 0.196 for prior, central4, central4+prior respectively). At T = 200 T = 200 T = 200 GeV, along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 only the two variants that include the coupling-prior forcing term cluster this tightly (u ( 10 − 3 , σ max ) = 0.599 , 0.598 u(10^{-3},\sigma_{\max}) = 0.599, 0.598 u ( 1 0 − 3 , σ m a x ) = 0.599 , 0.598 for prior and central4+prior), while the fourth-order stencil without the prior sits at a markedly different, intermediate value (u = 0.664 u = 0.664 u = 0.664 ) between the baseline (u = 0.71 u = 0.71 u = 0.71 ) and the prior-forced pair; along σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 , by contrast, all four variants including the baseline are within ∼ 20 % \sim20\% ∼ 20% of each other (u ( ρ max , 10 − 3 ) = 0.217 , 0.260 , 0.237 , 0.260 u(\rho_{\max},10^{-3}) = 0.217, 0.260, 0.237, 0.260 u ( ρ m a x , 1 0 − 3 ) = 0.217 , 0.260 , 0.237 , 0.260 for baseline, prior, central4, central4+prior respectively), so the sharp prior-versus-no-prior split seen along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 at T = 200 T = 200 T = 200 GeV does not carry over to the other slice. So stencil-order-insensitivity is not a fixed property of the setup: it holds at T = 100 T = 100 T = 100 GeV on both slices, and at T = 200 T = 200 T = 200 GeV only on the σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 slice, where it is instead the presence of the coupling-prior forcing term that separates the curves. Third, where a tight cluster does form, it is not uniformly close to the finite-temperature reference. At T = 200 T = 200 T = 200 GeV the prior-pair cluster overshoots the reference along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 by about 30% (u ≃ 0.60 u \simeq0.60 u ≃ 0.60 versus u = 0.46 u = 0.46 u = 0.46 ) but undershoots it along σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 by roughly a factor of two (u ≃ 0.26 u \simeq0.26 u ≃ 0.26 versus u = 0.60 u = 0.60 u = 0.60 ), essentially the same shortfall seen in the baseline there. At T = 100 T = 100 T = 100 GeV the cluster overshoots the reference along ρ = 10 − 3 \rho= 10^{-3} ρ = 1 0 − 3 by about 50% (u ≃ 0.37 u \simeq0.37 u ≃ 0.37 versus u = 0.25 u = 0.25 u = 0.25 ) and sits well below it along σ = 10 − 3 \sigma= 10^{-3} σ = 1 0 − 3 (u ≃ − 0.20 u \simeq-0.20 u ≃ − 0.20 versus a reference of u ≃ − 0.05 u \simeq-0.05 u ≃ − 0.05 ). So the direction of the mismatch (over- or under-shoot) and its size both vary with slice and temperature, rather than settling into a single clean pattern.
The true solution of the flow equation – the one that actually satisfies the Wetterich equation together with the correct initial and boundary conditions – would, as a consequence of the equation’s own physics, automatically be loosely consistent with the perturbative RGE running near the field-space origin; nothing needs to be imposed by hand for the exact trajectory to have this property. The coupling-prior penalty of § 4.5 was added precisely because the equation’s nonlinearity (the steep, poorly conditioned regime discussed above) lets the Newton–Krylov iterate drift away from that true trajectory while still keeping the pointwise residual R R R small, so a small ∥ R ∥ \lVert R\rVert ∥ R ∥ alone cannot certify that an iterate is still on-trajectory; loose perturbative consistency was meant to serve as an independent, cheap check on this drift, and the forcing term as a way to pull the iteration back when it is violated. Even so, Table 2 shows this check mostly failing to close: turning the prior on increases ∥ R ∥ \lVert R\rVert ∥ R ∥ in three of the four baseline/central4 pairs (from 0.144 to 0.217 and from 0.182 to 0.220 at T = 100 T = 100 T = 100 GeV; from 0.256 to 0.292 at T = 200 T = 200 T = 200 GeV for the central4 pair) rather than driving it toward zero. The one exception is the second-order pair at T = 200 T = 200 T = 200 GeV, where adding the prior actually lowers ∥ R ∥ \lVert R\rVert ∥ R ∥ (from 0.728 to 0.287) – but this is also the one case where the unforced baseline itself is the worst-converged of all eight runs, so the improvement reads less as the prior successfully enforcing consistency and more as a regression toward the typical ∥ R ∥ ∼ 0.2 \lVert R\rVert\sim0.2 ∥ R ∥ ∼ 0.2 –0.3 0.3 0.3 plateau seen elsewhere in the table. Either way, none of the eight runs comes close to the requested tolerance, and the penalty itself is not being satisfied at the stalled solution. In short, even this additional mechanism does not bring the drift under control, which is consistent with, and gives a mechanistic reason for, the slice-and temperature-dependent mismatch noted above. Taken together, these observations indicate that the addition of an external forcing term, rather than genuine convergence of the Newton–Krylov iteration, is what determines the fate of the stalled solution, which is the opposite of what a reliable solver should exhibit and is the practical motivation for the PINN-based construction of § 5 .
Network structure and loss functions
Network
The full potential is decomposed, in the spirit of the physics-motivated trial solutions of Ref. [34], into a physics-motivated baseline and a learnable residual:
u ( t , ρ , σ ) = u ~ t r e e ( t , ρ , σ ) + u ~ t h ( t , ρ , σ ) + N ( t , ρ , σ ) . (45) u(t,\rho,\sigma) = \widetilde{u}_{\mathrm{tree}}(t,\rho,\sigma) + \widetilde{u}_{\mathrm{th}}(t,\rho,\sigma) + N(t,\rho,\sigma). \tag*{(45)} u ( t , ρ , σ ) = u tree ( t , ρ , σ ) + u th ( t , ρ , σ ) + N ( t , ρ , σ ) . (45)
Here u ~ t r e e \widetilde{u}_{\mathrm{tree}} u tree is the running tree-level potential u Λ ∣ k U V → k u_\Lambda|_{k_{\mathrm{UV}}\to k} u Λ ∣ k UV → k with each of its five terms rescaled by a fixed O ( 1 ) O(1) O ( 1 ) factor (Appendix § 10 ), and u ~ t h = c T , ρ ρ τ 2 + c T , σ σ τ 2 \widetilde{u}_{\mathrm{th}} = c_{T,\rho}\rho\tau^2 + c_{T,\sigma}\sigma\tau^2 u th = c T , ρ ρ τ 2 + c T , σ σ τ 2 with ( c T , ρ , c T , σ ) = ( 0.2 , 0.1 ) (c_{T,\rho},c_{T,\sigma}) = (0.2, 0.1) ( c T , ρ , c T , σ ) = ( 0.2 , 0.1 ) is a crude leading high-temperature mass term. These two pieces are not meant to be accurate; they only absorb the bulk of the potential so that the network N N N is left to represent the smaller, nontrivial part of the flow. This is distinct from the analytic seed u s e e d u_{\mathrm{seed}} u seed of Eq. (50) , which is a genuine one-loop (ring-resummed) thermal potential used only as the fitting target for the network’s derivatives at the UV boundary t = 0 t = 0 t = 0 .
The neural network N N N is implemented via a modular architecture (Figure 4 ) that explicitly outputs not only the potential correction but also its first and second derivatives:
Figure 4. Schematic structure of the neural-network ansatz (BaseNet) used to represent the correction to the effective potential and its field derivatives. In the figure the subscripts ρ \rho ρ , σ \sigma σ are abbreviated r r r , s s s and the branch latent vectors are labelled z ρ z_\rho z ρ , z σ z_\sigma z σ . The three inputs ( t , ρ , σ ) (t,\rho,\sigma) ( t , ρ , σ ) are first processed by a shared stem MLP into a latent representation h h h . From this stem output, three branches are constructed. The u u u -branch applies an additional MLP to h h h and a linear head to produce the scalar output N N N , which corrects the potential itself. The ρ \rho ρ -branch takes the same stem output h h h , passes it through a dedicated MLP to obtain a latent vector z ρ z_\rho z ρ , and then uses two linear heads acting on z ρ z_\rho z ρ to generate the ρ \rho ρ -direction outputs N ρ N_\rho N ρ and N ρ ρ N_{\rho\rho} N ρρ . Similarly, the σ \sigma σ -branch processes h h h through another MLP to obtain z σ z_\sigma z σ , from which two linear heads produce N σ N_\sigma N σ and N σ σ N_{\sigma\sigma} N σ σ . The mixed branch receives as input the concatenation of the latent vectors z ρ z_\rho z ρ and z σ z_\sigma z σ together with the first-derivative outputs N ρ N_\rho N ρ and N σ N_\sigma N σ , but not the second derivatives N ρ ρ N_{\rho\rho} N ρρ or N σ σ N_{\sigma\sigma} N σ σ . This concatenated vector is passed through a final MLP and a linear head to yield the mixed second derivative N ρ σ N_{\rho\sigma} N ρ σ . Thus, the network outputs ( N , N ρ , N σ , N ρ ρ , N σ σ , N ρ σ ) (N,N_\rho,N_\sigma,N_{\rho\rho},N_{\sigma\sigma},N_{\rho\sigma}) ( N , N ρ , N σ , N ρρ , N σ σ , N ρ σ ) , with the mixed derivative N ρ σ N_{\rho\sigma} N ρ σ explicitly constructed from the field-direction latent features and first derivatives.
( N , N ρ , N σ , N ρ ρ , N σ σ , N ρ σ ) . (46) (N,N_\rho,N_\sigma,N_{\rho\rho},N_{\sigma\sigma},N_{\rho\sigma}). \tag*{(46)} ( N , N ρ , N σ , N ρρ , N σ σ , N ρ σ ) . (46)
We arrived at this design after an unsuccessful attempt at the more direct alternative: obtaining N ρ N_\rho N ρ , N σ N_\sigma N σ , N ρ ρ N_{\rho\rho} N ρρ , N σ σ N_{\sigma\sigma} N σ σ , N ρ σ N_{\rho\sigma} N ρ σ by nested automatic differentiation of a single-output network with respect to ρ , σ \rho,\sigma ρ , σ . Even combined with the same soft consistency penalty used throughout this paper (§ 5.2 ), this alternative still failed to converge to a physically sensible branch, in the same manner as the failure modes discussed in § 6.4 . Backpropagating the loss gradient through such a network requires third-order mixed field-and-weight derivatives – the loss depends on a second field derivative of the network, and training differentiates that dependence once more with respect to the weights – which measurably deepens the computational graph; we cannot, however, cleanly disentangle this added complexity from the same branch-selection non-uniqueness of the flow equation itself, since both were present in this comparison. This design avoids that repeated automatic differentiation and improves computational efficiency, at the cost of no longer guaranteeing that the derivative heads coincide with the true derivatives of N N N – addressed only approximately, via the self-consistency penalty of § 5.2 below.
The architecture consists of three components. First, a shared “stem” multilayer perceptron (MLP) processes the input ( t , ρ , σ ) (t,\rho,\sigma) ( t , ρ , σ ) into a latent representation, enabling efficient encoding of correlations among scale and field variables.
Second, separate branch networks are introduced for the ρ \rho ρ - and σ \sigma σ -directions. Each branch produces features that are mapped to first- and second-order derivatives (N ρ , N ρ ρ N_{\rho}, N_{\rho\rho} N ρ , N ρρ and N σ , N σ σ N_{\sigma}, N_{\sigma\sigma} N σ , N σ σ , respectively), reflecting the anisotropic role of the field directions in the RG flow.
Third, a mixed branch combines the latent representations from both field directions together with the first derivatives (N ρ , N σ N_{\rho}, N_{\sigma} N ρ , N σ ) to produce the mixed second derivative N ρ σ N_{\rho\sigma} N ρ σ . This construction allows the model to capture nontrivial correlations between ρ \rho ρ and σ \sigma σ .
All output heads are initialized to zero, ensuring that the initial network reproduces the baseline potential. Training then corresponds to learning deviations from the known tree-level and thermal structures, which stabilizes optimization and improves physical interpretability.
To enhance numerical performance, all inputs are linearly rescaled to the interval [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] . The corresponding Jacobian factors are consistently incorporated when relating derivatives in scaled and physical coordinates.
Concretely, the stem is a three-layer MLP of width 160; the u u u -branch is a single hidden layer of width 160, the ρ \rho ρ - and σ \sigma σ -branches two hidden layers of width 160 each, and the mixed branch two hidden layers of width 320 (its input being the concatenation z ρ ⊕ z σ ⊕ N ρ ⊕ N σ z_{\rho} \oplus z_{\sigma} \oplus N_{\rho} \oplus N_{\sigma} z ρ ⊕ z σ ⊕ N ρ ⊕ N σ , of dimension 322). All activations are SiLU and all output heads are zero-initialized linear maps, giving ≈ 3.9 × 10 5 \approx3.9 \times10^{5} ≈ 3.9 × 1 0 5 parameters per BaseNet. One such network is trained per domain block (§ 5.3 ), for a total of 10 networks along a one-dimensional slice and 50 for the full two-field problem. The remaining architectural and training hyperparameters are collected in Appendix § 10 .
This hybrid construction, combining analytic physics priors with a structured neural network ansatz, provides a flexible yet controlled representation of the effective potential suitable for functional renormalization group studies.
Loss functions
Collocation points ( t , ρ , σ ) (t,\rho,\sigma) ( t , ρ , σ ) within a given block are resampled at every training iteration from a mixture of 40% points drawn uniformly in the rescaled coordinates [ − 1 , 1 ] 3 [-1,1]^3 [ − 1 , 1 ] 3 and 60% points drawn as cos ( π U ( 0 , 1 ) ) \cos(\pi U(0,1)) cos ( π U ( 0 , 1 )) , which asymptotically reproduces the clustering of Chebyshev nodes near the block boundaries ± 1 \pm1 ± 1 and improves the resolution of the fit where the potential and its derivatives vary most rapidly. The network parameters are optimized by minimizing a weighted sum,
L = w p d e L p d e + w c o n s i s t L c o n s i s t + w b c L b c + w o v e r l a p L o v e r l a p + w s i g n L s i g n + w m a g L m a g , (47) \mathcal{L}=w_{\mathrm{pde}}\mathcal{L}_{\mathrm{pde}}+w_{\mathrm{consist}}\mathcal{L}_{\mathrm{consist}}+w_{\mathrm{bc}}\mathcal{L}_{\mathrm{bc}}+w_{\mathrm{overlap}}\mathcal{L}_{\mathrm{overlap}}+w_{\mathrm{sign}}\mathcal{L}_{\mathrm{sign}}+w_{\mathrm{mag}}\mathcal{L}_{\mathrm{mag}}, \tag*{(47)} L = w pde L pde + w consist L consist + w bc L bc + w overlap L overlap + w sign L sign + w mag L mag , (47)
whose six terms we describe in turn.
PDE residual. Let R ( t , ρ , σ ) R(t,\rho,\sigma) R ( t , ρ , σ ) denote the residual obtained by moving every term of the flow equation to one side. Only the RG-time derivative ∂ t u \partial_t u ∂ t u is evaluated by automatic differentiation of the scalar output; all field derivatives entering R R R (and hence the Goldstone and radial mass eigenvalues m ˉ G 2 \bar{m}^{2}_{G} m ˉ G 2 , m ˉ ± 2 \bar{m}^{2}_{\pm} m ˉ ± 2 ) are taken directly from the corresponding output heads described in Sec. 5.1, avoiding repeated automatic differentiation through the field directions. The PDE loss is then
L p d e = ⟨ R ( t , ρ , σ ) 2 ⟩ ( t , ρ , σ ) . (48) \mathcal{L}_{\mathrm{pde}}=\left\langle R(t,\rho,\sigma)^2\right\rangle_{(t,\rho,\sigma)} . \tag*{(48)} L pde = ⟨ R ( t , ρ , σ ) 2 ⟩ ( t , ρ , σ ) . (48)
For numerical stability, 1 + m ˉ G 2 1+\bar{m}_G^2 1 + m ˉ G 2 and 1 + m ˉ ± 2 1+\bar{m}_{\pm}^2 1 + m ˉ ± 2 are floored at 10 − 10 10^{-10} 1 0 − 10 before entering the threshold functions, the discriminant disc = ( M ˉ 11 2 − M ˉ 22 2 ) 2 + 4 ( M ˉ 12 2 ) 2 \operatorname{disc}=(\bar{M}_{11}^2-\bar{M}_{22}^2)^2+4(\bar{M}_{12}^2)^2 disc = ( M ˉ 11 2 − M ˉ 22 2 ) 2 + 4 ( M ˉ 12 2 ) 2 of the radial mass matrix [Eq. (21)] is floored at 10 − 6 10^{-6} 1 0 − 6 (so disc ≥ 10 − 3 \sqrt{\operatorname{disc}}\geq10^{-3} disc ≥ 1 0 − 3 ), and the resulting loop sum is clamped to [ − 2 , 2 ] [-2,2] [ − 2 , 2 ] . The comparatively large disc-floor is a training-stability measure rather than a physical regulator: although disc \sqrt{\operatorname{disc}} disc cancels from the loop sum itself (the two radial eigenvalues enter only through a combination symmetric under m ˉ + 2 ↔ m ˉ − 2 \bar{m}_{+}^2\leftrightarrow\bar{m}_{-}^2 m ˉ + 2 ↔ m ˉ − 2 , cf. Sec. 4.4), its gradient is back-propagated through the two eigenvalues along separate paths, and the near-degenerate limit disc → 0 \operatorname{disc}\to0 disc → 0 makes that gradient stiff. Occurrence of 1 + m ˉ G 2 < 0 1+\bar{m}_G^2<0 1 + m ˉ G 2 < 0 or 1 + m ˉ ± 2 < 0 1+\bar{m}_{\pm}^2<0 1 + m ˉ ± 2 < 0 after a majority of training has elapsed signals a pathological (tachyonic) configuration and aborts the run, which in practice never triggers once the boundary and consistency losses below have been imposed.
Consistency loss. Since u ρ u_\rho u ρ , u σ u_\sigma u σ , u ρ ρ u_{\rho\rho} u ρρ , u σ σ u_{\sigma\sigma} u σ σ , u ρ σ u_{\rho\sigma} u ρ σ are produced by dedicated output heads rather than obtained by differentiating u u u itself, nothing in the architecture guarantees that they coincide with the actual derivatives of the scalar output. We therefore add a self-consistency penalty comparing each head to a single automatic differentiation of the lower-order heads,
L c o n s i s t = ⟨ ( ∂ ρ u − u ρ ) 2 + w s 1 ( ∂ σ u − u σ ) 2 ⟩ + ⟨ ( ∂ ρ u ρ − u ρ ρ ) 2 + w s 2 ( ∂ σ u σ − u σ σ ) 2 + w r s [ ( ∂ ρ u σ − u ρ σ ) 2 + ( ∂ σ u ρ − u ρ σ ) 2 ] ⟩ . (49) \mathcal{L}_{\mathrm{consist}}=
\left\langle
(\partial_\rho u-u_\rho)^2+w_{s1}(\partial_\sigma u-u_\sigma)^2
\right\rangle
+\left\langle
(\partial_\rho u_\rho-u_{\rho\rho})^2+w_{s2}(\partial_\sigma u_\sigma-u_{\sigma\sigma})^2
+w_{rs}\left[(\partial_\rho u_\sigma-u_{\rho\sigma})^2+(\partial_\sigma u_\rho-u_{\rho\sigma})^2\right]
\right\rangle. \tag*{(49)} L consist = ⟨ ( ∂ ρ u − u ρ ) 2 + w s 1 ( ∂ σ u − u σ ) 2 ⟩ + ⟨ ( ∂ ρ u ρ − u ρρ ) 2 + w s 2 ( ∂ σ u σ − u σ σ ) 2 + w r s [ ( ∂ ρ u σ − u ρ σ ) 2 + ( ∂ σ u ρ − u ρ σ ) 2 ] ⟩ . (49)
evaluated at the same collocation points as L p d e \mathcal{L}_{\mathrm{pde}} L pde . The weights w s 1 = w s 2 = w r s = 5 w_{s1}=w_{s2}=w_{rs}=5 w s 1 = w s 2 = w r s = 5 used in the finite-temperature runs reported here upweight the singlet-direction and mixed terms relative to the Higgs-direction ones, whose gradients are typically larger simply because the Higgs direction carries the dominant tree-level curvature.
Boundary and inter-block continuity in t t t . At the UV edge t = 0 t=0 t = 0 of the outermost block in scale, the first derivatives of the network are fit to those of an analytic seed,
u s e e d ( t , ρ , σ ) = u t r e e ( t , ρ , σ ) + c c w u C W ( t , ρ , σ ) + 1 f i n i t e T u t h 1 -loop ( t , ρ , σ ) , (50) u_{\mathrm{seed}}(t,\rho,\sigma)=u_{\mathrm{tree}}(t,\rho,\sigma)+c_{\mathrm{cw}}u_{\mathrm{CW}}(t,\rho,\sigma)+\mathbb{1}_{\mathrm{finite}\ T}u_{\mathrm{th}}^{1\text{-loop}}(t,\rho,\sigma), \tag*{(50)} u seed ( t , ρ , σ ) = u tree ( t , ρ , σ ) + c cw u CW ( t , ρ , σ ) + 1 finite T u th 1 -loop ( t , ρ , σ ) , (50)
with c c w c_{\mathrm{cw}} c cw a switch for the one-loop Coleman–Weinberg piece (set to zero in the runs reported here, so that the network learns any zero-temperature loop correction purely from the flow equation itself) and u t h 1 -loop u_{\mathrm{th}}^{1\text{-loop}} u th 1 -loop the finite-temperature one-loop potential built from the same field content as the flow equation, evaluated with Debye-improved (ring-resummed) masses as discussed in Sec. 3.3. This gives L b c = ⟨ ( ∂ ρ u s e e d − u ρ ) 2 + ( ∂ σ u s e e d − u σ ) 2 ⟩ t = 0 \mathcal{L}_{\mathrm{bc}}=\langle(\partial_\rho u_{\mathrm{seed}}-u_\rho)^2+(\partial_\sigma u_{\mathrm{seed}}-u_\sigma)^2\rangle_{t=0} L bc = ⟨( ∂ ρ u seed − u ρ ) 2 + ( ∂ σ u seed − u σ ) 2 ⟩ t = 0 . For a block that is not the outermost one in t t t , this analytic anchor is replaced by matching against the already-trained, frozen neighboring block u p r e v u^{\mathrm{prev}} u prev evaluated at the shared time slice,
L b c = ⟨ ( u ρ − u ρ p r e v ) 2 + ( u σ − u σ p r e v ) 2 + w b c , 2 [ ( u ρ ρ − u ρ ρ p r e v ) 2 + ( u σ σ − u σ σ p r e v ) 2 + ( u ρ σ − u ρ σ p r e v ) 2 ] ⟩ , (51) \mathcal{L}_{\mathrm{bc}}=
\left\langle
(u_\rho-u_\rho^{\mathrm{prev}})^2+(u_\sigma-u_\sigma^{\mathrm{prev}})^2
+w_{\mathrm{bc},2}\left[(u_{\rho\rho}-u_{\rho\rho}^{\mathrm{prev}})^2+(u_{\sigma\sigma}-u_{\sigma\sigma}^{\mathrm{prev}})^2+(u_{\rho\sigma}-u_{\rho\sigma}^{\mathrm{prev}})^2\right]
\right\rangle, \tag*{(51)} L bc = ⟨ ( u ρ − u ρ prev ) 2 + ( u σ − u σ prev ) 2 + w bc , 2 [ ( u ρρ − u ρρ prev ) 2 + ( u σ σ − u σ σ prev ) 2 + ( u ρ σ − u ρ σ prev ) 2 ] ⟩ , (51)
enforcing (at least) C 1 C^1 C 1 continuity of the flowed potential across the interface; the weight of the second-derivative term is set to zero in the present runs.
Field-space interface matching. The same construction is applied across the boundaries between adjacent field-space blocks (Sec. 5.3): for two networks trained on overlapping ρ \rho ρ (and, in the two-dimensional setup, σ \sigma σ ) windows sharing a thin strip of width δ ρ = 0.02 \delta\rho=0.02 δ ρ = 0.02 , the first derivatives are required to agree throughout that strip,
L o v e r l a p = ⟨ ( u ρ l e f t − u ρ ) 2 + ( u σ l e f t − u σ ) 2 ⟩ ( t , ρ , σ ) ∈ o v e r l a p s t r i p . (52) \mathcal{L}_{\mathrm{overlap}}=
\left\langle
(u_\rho^{\mathrm{left}}-u_\rho)^2+(u_\sigma^{\mathrm{left}}-u_\sigma)^2
\right\rangle_{(t,\rho,\sigma)\in\ \mathrm{overlap\ strip}} . \tag*{(52)} L overlap = ⟨ ( u ρ left − u ρ ) 2 + ( u σ left − u σ ) 2 ⟩ ( t , ρ , σ ) ∈ overlap strip . (52)
where “left” denotes the previously trained, frozen neighboring block (or, in the full 2D grid, the frozen left and bottom neighbors in ρ \rho ρ and σ \sigma σ , respectively). This is the mechanism referred to in the Introduction as the derivative-matching condition tying the multi-domain decomposition together, in the spirit of the interface conditions of extended PINNs [35 ] .
Soft perturbative consistency (sign and magnitude penalties). The remaining two terms implement the soft consistency constraint mentioned in the abstract. On collocation points spanning the field-space domain, at the infrared edge of each RG-time block (t = − 1 t=-1 t = − 1 for the UV block, t = − 2 t=-2 t = − 2 for the IR block), we extract effective quartic couplings and quadratic mass terms directly from the network’s second- and first-derivative heads,
λ H N N ≡ 1 2 u ρ ρ , λ S N N ≡ 1 2 u σ σ , λ H S N N ≡ u ρ σ , (53) \lambda_H^{\mathrm{NN}} \equiv\frac{1}{2}u_{\rho\rho}, \qquad\lambda_S^{\mathrm{NN}} \equiv\frac{1}{2}u_{\sigma\sigma}, \qquad\lambda_{HS}^{\mathrm{NN}} \equiv u_{\rho\sigma}, \tag*{(53)} λ H NN ≡ 2 1 u ρρ , λ S NN ≡ 2 1 u σ σ , λ H S NN ≡ u ρ σ , (53)
a H N N ≡ u ρ − u ρ ρ ρ − u ρ σ σ , a S N N ≡ u σ − u σ σ σ − u ρ σ ρ . (54) a_H^{\mathrm{NN}} \equiv u_\rho-u_{\rho\rho}\rho-u_{\rho\sigma}\sigma, \qquad a_S^{\mathrm{NN}} \equiv u_\sigma-u_{\sigma\sigma}\sigma-u_{\rho\sigma}\rho. \tag*{(54)} a H NN ≡ u ρ − u ρρ ρ − u ρ σ σ , a S NN ≡ u σ − u σ σ σ − u ρ σ ρ . (54)
For each of these five combinations X ∈ { λ H , λ S , λ H S , a H , a S } X \in\{\lambda_H,\lambda_S,\lambda_{HS},a_H,a_S\} X ∈ { λ H , λ S , λ H S , a H , a S } we compare the network’s shift away from its UV tree-level value, Δ X N N ≡ X N N − X t r e e \Delta X_{\mathrm{NN}} \equiv X_{\mathrm{NN}}-X_{\mathrm{tree}} Δ X NN ≡ X NN − X tree , against the corresponding perturbative shift Δ X p e r t ( t ) \Delta X_{\mathrm{pert}}(t) Δ X pert ( t ) obtained from the two-loop running couplings (Appendix B), through two hinge-type penalties,
ℓ s i g n [ X ] = s β ( − sign ( Δ X p e r t ) Δ X N N ∣ X t r e e ∣ + margin ) , (55) \ell_{\mathrm{sign}}[X] = s_\beta\left(-\operatorname{sign}(\Delta X_{\mathrm{pert}})\frac{\Delta X_{\mathrm{NN}}}{\lvert X_{\mathrm{tree}}\rvert}+\operatorname{margin}\right), \tag*{(55)} ℓ sign [ X ] = s β ( − sign ( Δ X pert ) ∣ X tree ∣ Δ X NN + margin ) , (55)
ℓ m a g [ X ] = s β ( c − ∣ Δ X p e r t ∣ − ∣ Δ X N N ∣ ∣ X t r e e ∣ ) + s β ( ∣ Δ X N N ∣ − c + ∣ Δ X p e r t ∣ ∣ X t r e e ∣ ) , (56) \ell_{\mathrm{mag}}[X] = s_\beta\left(\frac{c_-\lvert\Delta X_{\mathrm{pert}}\rvert-\lvert\Delta X_{\mathrm{NN}}\rvert}{\lvert X_{\mathrm{tree}}\rvert}\right)+s_\beta\left(\frac{\lvert\Delta X_{\mathrm{NN}}\rvert-c_+\lvert\Delta X_{\mathrm{pert}}\rvert}{\lvert X_{\mathrm{tree}}\rvert}\right), \tag*{(56)} ℓ mag [ X ] = s β ( ∣ X tree ∣ c − ∣ Δ X pert ∣ − ∣ Δ X NN ∣ ) + s β ( ∣ X tree ∣ ∣ Δ X NN ∣ − c + ∣ Δ X pert ∣ ) , (56)
with ( c − , c + ) = ( 0.1 , 3.0 ) (c_-,c_+)=(0.1,3.0) ( c − , c + ) = ( 0.1 , 3.0 ) and s β s_\beta s β the same softplus hinge used for the grid-based prior in § 4.5 , here with the sharper β = 100 \beta=100 β = 100 . As in the grid case a smooth hinge rather than a hard ReLU is used deliberately: the ReLU corner at the constraint boundary was found to correlate with non-smooth, sign-oscillating behaviour in the learned curvature u ρ ρ u_{\rho\rho} u ρρ , and keeping the penalty differentiable across its activation threshold gives visibly cleaner second derivatives. L s i g n \mathcal{L}_{\mathrm{sign}} L sign and L m a g \mathcal{L}_{\mathrm{mag}} L mag are then obtained by averaging ℓ s i g n [ X ] \ell_{\mathrm{sign}}[X] ℓ sign [ X ] and ℓ m a g [ X ] \ell_{\mathrm{mag}}[X] ℓ mag [ X ] over X X X with relative weights of order unity (an order of magnitude smaller for λ H S \lambda_{HS} λ H S , whose perturbative shift is comparatively less reliable), restricted by a further relative-size mask that switches the penalty off wherever ∣ Δ X p e r t ∣ \lvert\Delta X_{\mathrm{pert}}\rvert ∣ Δ X pert ∣ itself is negligible compared to X t r e e X_{\mathrm{tree}} X tree . The magnitude window is intentionally loose: rather than forcing exact agreement with the perturbative result, which would defeat the purpose of a nonperturbative treatment, it merely forbids the trained network from developing quartic or mass corrections with the wrong sign or a wildly different order of magnitude from the (reliable, weak-coupling) perturbative expectation. In the finite-temperature runs the penalty is split by a smooth sigmoid mask at ρ = ρ c w , σ = σ c w \rho=\rho_{\mathrm{cw}},\sigma=\sigma_{\mathrm{cw}} ρ = ρ cw , σ = σ cw (Appendix A) into an inner and an outer field-space region. In the inner region, near the axes, the sign hinge is applied only to the singlet mass combination a S a_S a S and the magnitude band to a S a_S a S (and, at T = 100 G e V T=100\ \mathrm{GeV} T = 100 GeV , to a H a_H a H ), in each case with the analytic thermal shift of the mass combination subtracted first; the quartics there carry only the flat cap ∣ Δ λ i N N / λ i , t r e e ∣ ≤ 1 \lvert\Delta\lambda_i^{\mathrm{NN}}/\lambda_{i,\mathrm{tree}}\rvert\leq1 ∣ Δ λ i NN / λ i , tree ∣ ≤ 1 , since the nonanalytic thermal contributions near ρ = 0 \rho=0 ρ = 0 preclude a clean perturbative target for the curvature of u u u . In the outer region, where the thermal potential is smooth, the singlet quartic (and, at T = 100 G e V T=100\ \mathrm{GeV} T = 100 GeV , the Higgs quartic) instead carries a magnitude band around its two-loop RGE-running value. The Higgs mass combination a H a_H a H carries no sign constraint at finite temperature. This finite-temperature form is thus considerably more reduced than the grid-based coupling prior of § 4.5 to which it is otherwise analogous: the grid prior applies the sign hinge to all five combinations and uses the softer β = 10 \beta=10 β = 10 , whereas here only a S a_S a S is sign-constrained and β = 100 \beta=100 β = 100 . The full set of weight values is collected in Appendix A.
Soft anchoring of a learned solution to a cheap analytic model of this kind is used elsewhere in scientific machine learning – as multi-fidelity training of a correction network [36 ] , and, closest in spirit, as the enforcement of known exact conditions on machine-learned density functionals [37 ] . We arrived at the present construction independently; the difference is that in those settings the analytic guidance refines a model that already works, whereas here, as we discuss next, it turns out to be needed to select the physical branch at all.
In practice this constraint is not merely a convenience: switching it off (setting w s i g n = w m a g = 0 w_{\mathrm{sign}} = w_{\mathrm{mag}} = 0 w sign = w mag = 0 in Eq. (47) ) leaves the PDE, consistency, and boundary/interface losses alone unable to reliably pick out a physically sensible solution, since the flow equation’s nonlinearity admits nearby stationary points of that reduced loss with quartic or mass corrections of the wrong sign, or an unphysical order of magnitude, relative to the perturbative expectation near the field-space origin. Reaching the results reported below also required hand-tuning the relative weights of the combinations entering L s i g n \mathcal{L}_{\mathrm{sign}} L sign and L m a g \mathcal{L}_{\mathrm{mag}} L mag (Appendix § 10 ), and these weights were not identical between the two benchmark temperatures; we did not find a principled, weight-free way to impose this constraint. We report this openly as a practical limitation of the present setup rather than treat it as incidental, and return to it in § 6 .
Domain decomposition and multi-block training
A single network of the size described in § 5.1 cannot accurately represent the potential over the entire physically relevant range. The shape of u ( t , ρ , σ ) u(t,\rho,\sigma) u ( t , ρ , σ ) changes qualitatively between the UV, where it is close to the tree-level double expansion, and the IR, where it may develop a nontrivial multi-well structure; and fully-connected networks are additionally subject to spectral bias , fitting the low-frequency, large-scale part of a target function far faster under gradient descent than its fine-scale features [38 , 39 ] . In the PINN setting the same bias acts on the PDE residual and can be read off from the eigenspectrum of the associated neural tangent kernel [40 ] ; it worsens as the range of scales a single network must resolve grows – here, the nearly flat outer potential together with the sharp curvature that builds up near the emerging minima over the course of the flow. Restricting each subnetwork to a small sub-domain narrows this range and is a known mitigation [41 ] . We therefore partition the domain into a set of blocks and train one network per block, stitching the pieces together with the boundary and interface losses introduced above – an approach conceptually similar to the domain decomposition of extended PINNs [35 ] , applied here simultaneously in RG scale and in field space.
In the t t t direction, the full range t ∈ [ t r a n g e , 0 ] t \in[t_{\mathrm{range}},0] t ∈ [ t range , 0 ] (with t r a n g e = − 2 t_{\mathrm{range}} = -2 t range = − 2 in the runs reported here) is split into two contiguous blocks, t ∈ [ − 1 , 0 ] t \in[-1,0] t ∈ [ − 1 , 0 ] and t ∈ [ − 2 , − 1 ] t \in[-2,-1] t ∈ [ − 2 , − 1 ] , trained sequentially from the UV block to the IR block. The UV block is anchored to the analytic seed potential of Eq. (50) at t = 0 t = 0 t = 0 , while the IR block is instead matched, via L b c \mathcal{L}_{\mathrm{bc}} L bc , to the already-trained UV block at the shared slice t = − 1 t = -1 t = − 1 .
In the field directions, ρ ∈ [ 0 , ρ max ] \rho\in[0,\rho_{\max}] ρ ∈ [ 0 , ρ m a x ] (or σ ∈ [ 0 , σ max ] \sigma\in[0,\sigma_{\max}] σ ∈ [ 0 , σ m a x ] for the σ \sigma σ -mode runs, and both for the two-dimensional runs) with ρ max = σ max = 1.75 \rho_{\max} = \sigma_{\max} = 1.75 ρ m a x = σ m a x = 1.75 in units of k 2 k^2 k 2 is tiled by n b l o c k s = 5 n_{\mathrm{blocks}} = 5 n blocks = 5 windows on a uniform spacing Δ ρ = 0.35 \Delta\rho= 0.35 Δ ρ = 0.35 ; each window is widened by a thin band δ ρ = 0.02 \delta\rho= 0.02 δ ρ = 0.02 so that it overlaps its right neighbor, and the first derivatives are matched across that overlap by L o v e r l a p \mathcal{L}_{\mathrm{overlap}} L overlap . In the one-dimensional ρ \rho ρ - and σ \sigma σ -mode runs the network remains a genuine function u ( t , ρ , σ ) u(t,\rho,\sigma) u ( t , ρ , σ ) of both fields: only the tiled direction is not fixed to a point, while the other one is narrowed to [ 0 , 0.35 ] [0,0.35] [ 0 , 0.35 ] – wide enough to give the network real dependence on both fields, and hence a nontrivial mixed derivative u ρ σ u_{\rho\sigma} u ρ σ and consistency loss, without the cost of also tiling it into blocks. Critically, this still solves the same full two-field flow equation on a genuinely two-dimensional domain – both radial mass eigenvalues, the mixing term u ρ σ u_{\rho\sigma} u ρ σ , and the network’s mixed-derivative head are all active – so the ρ \rho ρ - and σ \sigma σ -mode runs are not a simplified single-field problem; only the domain and the block tiling are restricted. The blocks are trained sequentially, each one referencing its already-trained, frozen left neighbor in ρ \rho ρ (and, for the full 2D grid, also its bottom neighbor in σ \sigma σ , the two field directions being swept as an outer/inner loop over a 5 × 5 5 \times5 5 × 5 grid of windows). Since every field-space window is itself further divided into the two t t t -sub-blocks described above, the potential over the full physically relevant domain is ultimately represented by a patchwork of up to 5 × 5 × 2 = 50 5 \times5 \times2 = 50 5 × 5 × 2 = 50 independently trained BaseNet instances for the two-field problem (or 5 × 2 = 10 5 \times2 = 10 5 × 2 = 10 along a one-dimensional slice), each one accurate only within, and continuously matched across, its own sub-domain.
Optimizer and training schedule
Each block is trained with the SOAP optimizer [42 ] , which combines an Adam-like update rule [43 ] with a Shampoo-style preconditioner [44 ] , together with a cosine-annealed learning-rate schedule [45 ] decaying from a base learning rate 3 × 10 − 4 3 \times10^{-4} 3 × 1 0 − 4 to η min = 10 − 5 \eta_{\min} = 10^{-5} η m i n = 1 0 − 5 over 10 4 10^4 1 0 4 iterations per block. At each iteration the losses are evaluated on freshly resampled collocation points – 7168 for the PDE residual, 4096 for the boundary condition, 2048 for the field-space interface, and 1024 for the perturbative-consistency penalty – rather than on a static grid. Once the main schedule completes, additional short refinement rounds of 100 iterations each are carried out at the final, frozen learning rate – without further decay and without intermediate checkpointing – for up to 25 such rounds, stopping early as soon as the total loss falls below a fixed threshold (10 − 9 10^{-9} 1 0 − 9 ). Between successive rounds the acceptance band of the magnitude penalty L m a g \mathcal{L}_{\mathrm{mag}} L mag is widened by a fixed geometric factor, so that this deliberately loose constraint does not obstruct the final convergence of the PDE and boundary terms. This low-noise fine-tuning stage improves the accuracy of the derivative outputs that the downstream analysis of § 3 (vacuum locations, bounce action) relies on, without reintroducing the risk of instability that a further scheduled decay of the learning rate could otherwise carry.
Training uses PyTorch on a single NVIDIA RTX A6000 GPU; at ≈ 0.2 \approx0.2 ≈ 0.2 s per iteration, a full ρ \rho ρ - or σ \sigma σ -direction run (10 blocks) takes up to ≈ 7 \approx7 ≈ 7 h including refinement rounds, and the two-dimensional run (50 blocks) up to ≈ 37 \approx37 ≈ 37 h – consistent with the checkpoint/auto-resume design needed to span multiple Paperspace sessions.
Figure 5 shows a representative example of the resulting per-block training curves, for the five ρ \rho ρ -direction blocks at T = 200 T = 200 T = 200 GeV (the σ \sigma σ -direction and the other temperature behave similarly and are omitted here). The PDE and boundary-condition losses of each block fall by two to four orders of magnitude over the course of training. The occasional spikes visible for the outer blocks (i ρ = 2 i_\rho= 2 i ρ = 2 –4) fall at the boundaries between the short refinement rounds, where the widening of the L m a g \mathcal{L}_{\mathrm{mag}} L mag acceptance band shifts the penalty landscape discontinuously; the learning rate itself is held fixed across these rounds. After each spike the loss relaxes back down to, or below, its pre-spike level.
Figure 5. Training-loss histories (PDE residual and boundary/interface condition terms, Eq. (47) ) for the five ρ \rho ρ -direction field-space blocks of § 5.3 , at T = 200 T = 200 T = 200 GeV, on a log scale. Block 1 covers the UV half of the RG-time range t ∈ [ − 1 , 0 ] t \in[-1,0] t ∈ [ − 1 , 0 ] and block 2 the IR half t ∈ [ − 2 , − 1 ] t \in[-2,-1] t ∈ [ − 2 , − 1 ] .
Results
A first check: the Higgs-only reduction (PINN)
We begin, as with the grid-based solver in § 4.3 , with the simplest version of the problem: the singlet-free reduction of the model – a single field direction ρ \rho ρ , no radial mass matrix and no disc \sqrt{\operatorname{disc}} disc anywhere. This isolates the role of the soft perturbative-consistency penalty of § 5.2 , which was motivated in part by the singlet-sector eigenvalue mixing, from any singlet-specific structure: if the network still depends on the penalty in this reduction, that dependence cannot be an artifact of the 2 × 2 2 \times2 2 × 2 mass matrix. We retrained the network for the reduced model at T = 200 T = 200 T = 200 GeV, with and without the penalty (w s i g n = w m a g = 0 w_{\mathrm{sign}} = w_{\mathrm{mag}} = 0 w sign = w mag = 0 in Eq. (47) ), keeping the architecture, the multi-block decomposition, the training schedule and the collocation sampling unchanged.
The result is shown in Figure 6 , at the UV boundary t = 0 t = 0 t = 0 and the IR end of the flow t = − 2 t = -2 t = − 2 . With the penalty, the network tracks the ring-resummed finite-temperature reference closely: at t = − 2 t = -2 t = − 2 it agrees with the reference to within ∼ 3 % \sim3\% ∼ 3% over the full field range (u = 0.53 u = 0.53 u = 0.53 versus 0.55 0.55 0.55 at ρ = ρ max = 1.75 \rho= \rho_{\max} = 1.75 ρ = ρ m a x = 1.75 , and similarly at every intermediate ρ \rho ρ ). Without the penalty, training on the PDE, consistency and boundary/interface losses alone collapses onto a spurious, nearly flat solution – u ≃ 0 u \simeq0 u ≃ 0 across the domain at t = − 2 t = -2 t = − 2 , even dipping slightly negative at intermediate ρ \rho ρ – that matches neither the finite-temperature reference nor the zero-temperature Coleman–Weinberg curve. The two solutions are essentially indistinguishable at the UV boundary t = 0 t = 0 t = 0 and diverge progressively as the flow is integrated toward the IR.
Figure 6. First check of the PINN on the Higgs-only reduction at T = 200 T = 200 T = 200 GeV: the stitched network prediction (with the gauge-sector ring correction of § 3.3 added back) along ρ \rho ρ at the UV boundary t = 0 t = 0 t = 0 (left) and the IR end of the flow t = − 2 t = -2 t = − 2 (right), trained with the soft perturbative-consistency penalty of § 5.2 (red) and without it, w s i g n = w m a g = 0 w_{\mathrm{sign}} = w_{\mathrm{mag}} = 0 w sign = w mag = 0 (blue). The penalized network follows the RGE-improved, Arnold–Espinosa ring-resummed finite-temperature reference (green dotted) throughout the flow; the unpenalized network drifts to a spurious nearly-flat solution that tracks neither the finite-temperature nor the zero-temperature Coleman–Weinberg reference (grey dash-dotted).
Since this reduced system has none of the singlet mass-matrix structure, the failure of the unconstrained network to find the physical branch, and the effectiveness of the loose penalty in restoring it, cannot be attributed to the eigenvalue-crossing non-analyticity. Both are properties of the nonlinear loop-plus-thermal flow equation itself and are already present with a single field: the obstruction is a steep-but-smooth regime of the equation, amplified toward the IR by the − 4 u -4u − 4 u term and the dimensionless rescaling, rather than a genuine singularity. The full-system slices of § 6.2 are read the same way – the residual deviations there, along both field directions, reflect the accuracy a PINN of this size and training budget reaches on an equation this nonlinear and carry no signature of the eigenvalue-crossing non-analyticity. The grid-based solver behaves the same way in this reduction (§ 4.3 ): without the coupling prior it produces an untrustworthy solution, and the prior helps but, unlike here, does not bring it to the finite-temperature answer – a contrast we return to in § 6.4 .
Slice comparison with the perturbative and relaxation benchmarks
As a check of the trained network against the same physics used to benchmark the grid-based solver in § 4 , Figure 7 and Figure 8 show slices of the PINN’s potential u ( t , ρ , σ ) u(t,\rho,\sigma) u ( t , ρ , σ ) along the σ \sigma σ -direction (ρ = 0 \rho= 0 ρ = 0 fixed) and the ρ \rho ρ -direction (σ = 0 \sigma= 0 σ = 0 fixed) – the true field-space origin, since the multi-block training of § 5.3 covers the full domain down to ρ = σ = 0 \rho= \sigma= 0 ρ = σ = 0 , unlike the domain-restricted relaxation solver of § 4.4 – at the two RG-time endpoints t = 0 t = 0 t = 0 (the UV boundary, where the network is anchored to the analytic seed potential of Eq. (50) ) and t = t r a n g e = − 2 t = t_{\mathrm{range}} = -2 t = t range = − 2 (the IR end of the flow), for the same two thermal benchmark points T = 100 T = 100 T = 100 and T = 200 T = 200 T = 200 GeV used in § 4.6 . Each panel compares four curves: the exact tree-level potential (black), the network’s raw prediction stitched continuously across the five field-space training blocks of § 5.3 , with the piecewise per-block shifts of § 5.3 fixed by continuity and an overall additive constant fixed by u ( t , 0 , 0 ) = 0 u(t,0,0) = 0 u ( t , 0 , 0 ) = 0 (red dashed, “NN+ring stitched”), a reference built from the tree-level potential with RG-improved (two-loop running) couplings, augmented only at t = t r a n g e t = t_{\mathrm{range}} t = t range by a static one-loop Coleman–Weinberg correction (blue dash-dotted, “RGE (T = 0 T = 0 T = 0 )”), and the same RG-improved tree supplemented instead by the Arnold–Espinosa-resummed finite-temperature one-loop potential (green dotted, “finite T”) – the closest analogue, among the reference curves available here, to the nonperturbative trajectory the network is trying to reproduce. Following the argument of § 3.3 , only the external gauge sector’s ring (Debye-daisy) correction is added back explicitly on top of the raw network output to form the “NN+ring” curve, since the scalar sector’s own ring resummation is already generated internally, nonperturbatively, by the FRG flow itself; the “finite T” reference, by contrast, being built entirely from perturbation theory, must include the scalar Goldstone/radial ring corrections by hand as well, in addition to the gauge ones.
Figure 7. PINN prediction along the ρ \rho ρ -direction (σ = 0 \sigma= 0 σ = 0 fixed, top) and the σ \sigma σ -direction (ρ = 0 \rho= 0 ρ = 0 fixed, bottom) at T = 200 T = 200 T = 200 GeV. Each panel shows the UV boundary t = 0 t = 0 t = 0 (left) and the IR end of the flow t = − 2 t = -2 t = − 2 (right): tree-level potential (black), the stitched network prediction with the gauge-sector ring correction of § 3.3 added back (red dashed, “NN+ring stitched”), the RGE-improved zero-temperature reference (blue dash-dotted), and the same reference with the finite-temperature piece added (green dotted). Vertical dotted lines mark the boundaries between the five field-space training blocks of § 5.3 .
Figure 8. (a) ρ \rho ρ -direction. (b) σ \sigma σ -direction. Same as Figure 7 , at T = 100 T = 100 T = 100 GeV.
At t = 0 t = 0 t = 0 all curves coincide with the tree-level potential essentially exactly by construction, since the UV boundary condition pins the network (and the perturbative references) to the same seed. The nontrivial comparison is therefore at t = − 2 t = -2 t = − 2 . At T = 100 T = 100 T = 100 GeV the network reproduces the finite-temperature reference closely along both field-space axes: the RMS deviation across the slice is below 1% of the reference’s variation along the singlet (σ \sigma σ ) direction and about 3% along the Higgs (ρ \rho ρ ) direction (at the field-space maxima, u = 0.28 u = 0.28 u = 0.28 against 0.27 and u = − 0.10 u = -0.10 u = − 0.10 against − 0.09 -0.09 − 0.09 respectively, the latter pair both small against the tree value − 0.35 -0.35 − 0.35 there). At T = 200 T = 200 T = 200 GeV the agreement loosens to an RMS deviation of 5–6% of the slice range in both directions, with the network running below the reference through the interior of each slice (crossing it near the outer edge along ρ \rho ρ ); at the field maxima u = 0.60 u = 0.60 u = 0.60 against 0.57 along ρ \rho ρ and u = 0.43 u = 0.43 u = 0.43 against 0.48 along σ \sigma σ . Here the σ \sigma σ -direction is the singlet direction (Higgs field held at zero) and the ρ \rho ρ -direction the Higgs direction.
The singlet direction is reproduced somewhat more accurately than the Higgs direction, at both temperatures and by the grid-based relaxation solver of § 4.6 as well. We do not read anything structural into this: the two field directions are inequivalent already at the Lagrangian level – the Higgs carries the gauge and top-Yukawa loops and a different quartic and mass parameter from the singlet – so there is no symmetry between them for the numerics to respect, and a difference in how accurately each is resolved is unremarkable. It is in particular not a signature of the disc → 0 \to0 → 0 eigenvalue-crossing non-analyticity of § 4.4 : the collocation points for these blocks stay well away from any 1 + m ˉ 2 = 0 1 + \bar{m}^{2} = 0 1 + m ˉ 2 = 0 threshold, restricting the field range away from disc = 0 = 0 = 0 leaves the ρ \rho ρ -direction offset essentially unchanged, and the Higgs-only reduction of § 6.1 shows the same residual difficulty with the singlet, and d i s c \sqrt{\mathrm{disc}} disc , removed entirely.
The residual deviation from the finite-temperature reference, small at T = 100 T = 100 T = 100 and growing to a few percent at T = 200 T = 200 T = 200 , is best understood simply as the accuracy a PINN of this size and training budget attains on a flow equation this strongly nonlinear – of a piece with the network’s documented need for the soft consistency penalty and hand-tuned loss weights (§ 5.2 , § 6.4 ), and with the stall of the Newton–Krylov iteration on the grid-based side (§ 4.4 ). It does not call for a separate physical or analytic explanation. We have not carried out a systematic scan in network width or training budget to confirm this attribution, which we leave to future work. Across the block interfaces the matched first derivative (u ρ u_\rho u ρ in the ρ \rho ρ direction, u σ u_\sigma u σ on the singlet slices) agrees between neighboring blocks to about 2% of the local slope in the median and to under ∼ 10 % \sim10\% ∼ 10% at the ninetieth percentile – a few × 10 − 3 \times10^{-3} × 1 0 − 3 in absolute terms, comparable to the pointwise deviations from the reference quoted above – at both benchmark temperatures, with larger relative values confined to the immediate neighborhood of the origin at the deepest infrared times, where the slope itself passes through zero. The interface conditions thus enforce C 1 C^1 C 1 continuity to within the overall solution accuracy rather than exactly.
The derivative heads themselves, tied to the true derivatives of the scalar output only softly through L c o n s i s t \mathcal{L}_{\mathrm{consist}} L consist (§ 5.2 ), stay close to them in the converged networks: sampled over the collocation domain, the RMS relative mismatch ⟨ ( ∂ ρ u − u ρ ) 2 ⟩ 1 / 2 / ⟨ u ρ 2 ⟩ 1 / 2 \langle(\partial_\rho u-u_\rho)^2\rangle^{1/2}/\langle u_\rho^2\rangle^{1/2} ⟨( ∂ ρ u − u ρ ) 2 ⟩ 1/2 / ⟨ u ρ 2 ⟩ 1/2 between each head and one automatic differentiation of the next-lower head is below 0.4% for every first- and second-derivative head, and under 0.7% in the worst individual block, at both benchmark temperatures – several times tighter than the median interface C 1 C^1 C 1 mismatch and well below the deviation from the resummed-perturbation reference.
Toward the two-dimensional potential
The results above are all restricted to one-dimensional slices in field space – the axes σ = 0 \sigma= 0 σ = 0 and ρ = 0 \rho= 0 ρ = 0 of the full system (§ 6.2 ), and the singlet-free reduction of § 6.1 – which is sufficient to fix the location and depth of each vacuum separately but does not resolve the potential barrier between them – the quantity that ultimately controls whether the electroweak step of a two-step transition (§ 3.1 ) is strongly first-order. Assessing that barrier requires the full two-dimensional potential u ( t , ρ , σ ) u(t, \rho, \sigma) u ( t , ρ , σ ) , and as a first step in this direction we have extended the multi-block architecture of § 5.3 from the two one-dimensional slices to a genuine 5 × 5 5 \times5 5 × 5 grid of blocks in ( ρ , σ ) (\rho, \sigma) ( ρ , σ ) , stitched together with the same continuity matching used in one dimension. Training of this 2D network is at present only preliminary and has not been tuned to the same accuracy as the one-dimensional results reported above; the example shown here should therefore be read as a demonstration of the framework rather than as a converged physics result. The harder convergence of the two-dimensional run is therefore not a qualitative jump to a more complex equation – it is the same flow equation and architecture already exercised by the one-dimensional runs (§ 5.3 ), now tiled over the full domain in both directions (5 × 5 5 \times5 5 × 5 instead of 5 × 1 5 \times1 5 × 1 blocks) rather than a narrow slab in the secondary field, and consequently exposed to the genuinely mixed, far-from-axis region that the one-dimensional runs never had to resolve.
Figure 9 shows the potential along the path-deformed O ( 3 ) O(3) O ( 3 ) bounce trajectory connecting the two axis vacua ( h , s ) = ( 0 , v S ) (h, s) = (0, v_S) ( h , s ) = ( 0 , v S ) and ( h , s ) = ( v , 0 ) (h, s) = (v, 0) ( h , s ) = ( v , 0 ) at t = − 2 t = -2 t = − 2 and T = 100 T = 100 T = 100 GeV – that is, the same trajectory obtained by minimizing the bounce action below (§ 3.2 , via the deformation of Ref. [26 ] ), rather than a straight line imposed by hand. The “NN+ring” and “finite T T T ” curves are each drawn along their own bounce trajectory in the respective potential, closed onto that potential’s own pair of axis vacua (both located from the corresponding 2D prediction); the potential accordingly approaches each endpoint quadratically, with vanishing slope, although the flat region is narrow on the scale of the plot because the wells are steep in the physical amplitudes ( h , s ) (h, s) ( h , s ) . The resulting curve already reproduces the qualitative shape expected for a two-step history – a single potential barrier separating the singlet-broken and electroweak minima – although the height and asymmetry of the barrier still differ visibly from the finite-temperature reference (the network’s barrier, u max − u f a l s e ≈ 0.017 u_{\max} - u_{\mathrm{false}} \approx0.017 u m a x − u false ≈ 0.017 , is about half the reference’s ≈ 0.032 \approx0.032 ≈ 0.032 ), consistent with the preliminary state of the training noted above.
With a barrier resolved, the framework already delivers a nucleation-scale observable end to end. Taking the flowed two-dimensional potential of Figure 9 as input, we solve the O ( 3 ) O(3) O ( 3 ) bounce equation of § 3.2 along the path-deformed trajectory [26 ] connecting the two vacua, with the potential and its gradient supplied analytically by the trained network and the kinetic terms taken canonically normalized (Z H = Z S = 1 Z_H = Z_S = 1 Z H = Z S = 1 ) in the bounce functional, i.e. dropping the wave-function-renormalization running in this final step. Dropping the η \eta η running here is well justified: η H ≃ 2 16 π 2 ( 3 y t 2 − 9 4 g 2 2 − 3 4 g 1 2 ) ≈ 0.02 \eta_H \simeq\frac{2}{16\pi^2}\left(3y_t^2 - \frac{9}{4}g_2^2 - \frac{3}{4}g_1^2\right) \approx0.02 η H ≃ 16 π 2 2 ( 3 y t 2 − 4 9 g 2 2 − 4 3 g 1 2 ) ≈ 0.02 so Z H Z_H Z H varies by only a few percent over ∣ t ∣ ≤ 2 |t| \le2 ∣ t ∣ ≤ 2 , η S ∼ 10 − 4 \eta_S \sim10^{-4} η S ∼ 1 0 − 4 is negligible, and the resulting shift in S 3 / T S_3/T S 3 / T is far below the barrier-convergence uncertainty that dominates the quoted number. This gives S 3 ( T ) / T ≈ 3 × 10 2 S_3(T)/T \approx3 \times10^2 S 3 ( T ) / T ≈ 3 × 1 0 2 at T = 100 T = 100 T = 100 GeV, against S 3 ( T ) / T ≈ 6 × 10 2 S_3(T)/T \approx6 \times10^2 S 3 ( T ) / T ≈ 6 × 1 0 2 obtained from the perturbative Arnold–Espinosa finite-temperature reference potential evaluated the same way – a factor of two below the reference, consistent with the still-visible barrier-height difference between the “NN+ring” and “finite T T T ” curves of Figure 9 . Both values sit well above the nucleation criterion S 3 ( T n ) / T n ≃ 140 S_3(T_n)/T_n \simeq140 S 3 ( T n ) / T n ≃ 140 of § 3.2 , so this benchmark places T n T_n T n below 100 GeV; the un-converged barrier and the absence of a temperature scan preclude a quantitative determination of T n T_n T n itself. A straight-line single-field instanton bound on S 3 S_3 S 3 , attempted as a cross-check, did not converge numerically for this network snapshot and is not quoted. The remaining gap is dominated by the still-incomplete convergence of the 2D barrier rather than by the bounce computation; with the present training we cannot separate this convergence error from any genuine nonperturbative reduction of the barrier along the flow. We therefore quote the number as an order-of-magnitude, end-to-end demonstration rather than a converged transition strength. Bringing the 2D network to the same level of convergence as the 1D slices, propagating the residual k I R k_{\mathrm{IR}} k IR -truncation systematic, and turning this into a determination of T n T_n T n and ξ c \xi_c ξ c , is left for future work.
Figure 9. Preliminary two-dimensional PINN prediction (with the gauge-sector ring correction of § 3.3 added back, “NN+ring”) for the potential along the path-deformed O ( 3 ) O(3) O ( 3 ) bounce trajectory connecting the singlet-only vacuum ( h , s ) = ( 0 , v S ) (h,s)=(0,v_S) ( h , s ) = ( 0 , v S ) and the electroweak vacuum ( h , s ) = ( v , 0 ) (h,s)=(v,0) ( h , s ) = ( v , 0 ) , at t = − 2 t=-2 t = − 2 and T = 100 T=100 T = 100 GeV, parametrized by normalized arc length ℓ \ell ℓ , with ℓ = 0 \ell=0 ℓ = 0 at the singlet-only vacuum and ℓ = 1 \ell=1 ℓ = 1 at the electroweak vacuum. The “NN+ring” and “finite T T T ” curves are each drawn along the deformed bounce trajectory in their own potential, connecting that potential’s own pair of axis vacua (both located from the corresponding 2D prediction; these are the same trajectories entering the S 3 S_3 S 3 computation in the text). The 2D training underlying this figure is not yet converged to the same accuracy as the one-dimensional results above; see the discussion in the text.
Practical role and limitations of the soft consistency penalty
As anticipated in § 5.2 , the soft perturbative-consistency penalty (L s i g n \mathcal{L}_{\rm sign} L sign and L m a g \mathcal{L}_{\rm mag} L mag ) turned out to be necessary rather than optional: without it, training on the PDE residual, consistency, and boundary/interface losses alone did not reliably converge to a physically sensible solution, and reaching the results above also required hand-tuning the relative weights of L s i g n \mathcal{L}_{\rm sign} L sign and L m a g \mathcal{L}_{\rm mag} L mag across the constrained combinations (Appendix § 10 ), with weights that differed between the two benchmark temperatures. We report this openly as a practical limitation of the present setup.
What the penalty actually enforces is weak. It fixes only the sign and the order of magnitude of a few local low-order coefficients of the potential: the penalized mass combination – the singlet a S a_S a S , and at T = 100 T=100 T = 100 GeV also the Higgs a H a_H a H – has its shift away from tree level held between roughly a tenth and three times the corresponding analytic (RGE-running plus thermal) shift, and the quartic curvatures are held within about a factor of two of their tree-level (inner region) or RGE-running (outer region) value. Nothing is imposed on the height or position of the barrier, on the location or depth of the infrared minima, or on the potential away from the two RG-time slices t = − 1 , − 2 t=-1,-2 t = − 1 , − 2 on which the penalty acts. In the ideal case – a weighting on which the converged solution does not depend – the penalty would then act purely as a branch selector: of the several nearby near-stationary points that the reduced loss admits for this nonlinear flow (§ 5.2 ), it would pick out the one consistent with perturbation theory in sign and magnitude, without dictating the result, so that the reported potential remains a solution of the flow equation, simply the physically sensible one. The limitation documented below is that this ideal is not quite reached in practice: the converged infrared potential keeps a residual, quantitatively relevant dependence on the hand-tuned weights and on the analytic thermal target.
It is worth being precise about where this difficulty actually sits. Both methods handle the tree-level problem without any such guidance: the tree-only relaxation system is linear and solved essentially exactly by a single LU factorization (§ 4.2 ), and the PINN likewise reproduces the tree-level potential to high accuracy at the UV boundary t = 0 t = 0 t = 0 by construction (see above). The breakdown in both cases is specifically tied to switching on the loop contributions – the E e x t E_{\mathrm{ext}} E ext term and the coth threshold functions of Eq. (17) , including their finite-temperature part – which is what makes the flow equation genuinely nonlinear and is exactly where the Newton–Krylov iteration stalls (§ 4.4 ) and where the PINN requires the soft consistency penalty to stay on a physically sensible branch. The comparison between the two methods should therefore be read as a comparison of how each one copes once loop (including thermal) effects are included, not of the tree-level problem, which both solve comfortably on their own.
At the same time, this is a meaningful contrast with, rather than a repetition of, the outcome of § 4 : there, an essentially analogous coupling-prior forcing term was added directly to the Newton–Krylov residual (§ 4.5 ), and it did not rescue the solver – turning it on drove the final residual norm up rather than down in three of the four solver/stencil pairs (Table 2 ), and in the one pair where it helped only by regressing the worst-converged run toward the same plateau as the rest, so the grid-based method could not be brought to a physically sensible solution either with or without this guidance. For the PINN, by contrast, adding the same kind of loose perturbative guidance, once its relative weights were tuned, did yield the physically sensible, closely tree/finite-temperature-tracking solutions reported above. So while neither method solves the flow equation from the PDE residual and boundary/interface conditions alone, only the PINN was able to make productive use of the additional perturbative guidance once supplied; the rigidity of the fixed finite-difference stencil, rather than the idea of a coupling prior itself, appears to be what prevented the relaxation solver from benefiting in the same way.
We regard removing the PINN’s remaining dependence on hand-tuned weights, whether through a better-conditioned network architecture, a more informative choice of collocation points, or a principled rather than hand-tuned weighting of the consistency terms, as an important open problem rather than a solved part of the present framework.
The dependence runs deeper than the choice of weights. In the finite-temperature runs the mass-like combinations a H , a S a_H, a_S a H , a S are compared against an analytic thermal target obtained by differentiating the one-loop ring-resummed thermal potential (§ 5.2 ), and we find the converged infrared potential to retain a visible sensitivity both to that target and to which combinations receive a penalty at all – beyond the sensitivity to the weights and to the on/off switch already discussed. The agreement of the mass sector with the perturbative finite-temperature reference is therefore in part imposed through this target rather than a pure output of the flow. The genuinely unconstrained, nonperturbative content of the solution sits in the quartic/shape sector, which carries only the loose cap ∣ Δ λ i N N / λ i , t r e e ∣ ≲ 1 \lvert\Delta\lambda_i^{\mathrm{NN}}/\lambda_{i,\mathrm{tree}}\rvert\lesssim1 ∣ Δ λ i NN / λ i , tree ∣ ≲ 1 – and this is precisely the sector that is not yet robust against the above choices. Making the flow equation together with the genuine UV and thermal boundary data fix the mass sector on their own, without an analytic thermal target, is in our view the central open problem, ahead of raw network capacity.
Conclusions and discussion
We have applied physics-informed neural networks to the Wetterich flow equation of a Z 2 \mathbb{Z}_2 Z 2 -symmetric singlet extension of the Standard Model, including SM gauge/Yukawa loops and finite-temperature effects, and used the resulting effective potential to demonstrate an end-to-end electroweak-phase-transition pipeline. A mesh-free, multi-block PINN architecture reproduces the tree-level potential essentially exactly at the UV boundary by construction, and at the IR end of the flow tracks the perturbative (RGE-improved, finite-temperature) reference to within a few percent along both field-space axes – most closely at T = 100 T = 100 T = 100 GeV, with the deviation growing to 5–6% at T = 200 T = 200 T = 200 GeV, at the level of accuracy expected of a PINN on an equation this nonlinear (§ 6 ). As a preliminary step toward the two-dimensional potential needed to assess the barrier between the singlet-broken and electroweak minima, we also trained a 5 × 5 5 \times5 5 × 5 -block extension of the network and used it to locate both vacua directly and evaluate the potential along the line connecting them; while not yet converged to the same accuracy as the one-dimensional results, it already reproduces the qualitative barrier structure expected of a two-step transition.
To assess whether this outcome required the flexibility of a neural network in particular, or would have followed from discretizing the same flow equation by more conventional means, we benchmarked the PINN against a grid-based Newton–Krylov relaxation solver applied to the identical equation (§ 4 ). The comparison is informative precisely because it is asymmetric: the relaxation solver does not converge to a physically sensible solution once loop and thermal effects are switched on, and adding a coupling-prior forcing term analogous to the PINN’s soft perturbative-consistency penalty does not rescue it – in most of the runs it drives the residual up rather than down (Table 2 ). The PINN, by contrast, does reach physically sensible solutions once the analogous penalty is included, though only after hand-tuning its relative weights. The rigidity of the fixed finite-difference stencil, rather than the coupling-prior idea itself, appears to be what prevents the grid method from benefiting from the same guidance (§ 6.4 ).
Repeating both calculations for a singlet-free reduction of the model – Higgs plus the external gauge and top sector only, with no 2 × 2 2 \times2 2 × 2 radial mass matrix and hence none of the near-degenerate eigenvalue mixing of the full system – shows that this contrast does not originate in the singlet sector (§ 4.3 , § 6.1 ). In the reduction the grid solver behaves better – it can even reach tolerance – but still converges to a spurious rather than the physical solution, while the unconstrained PINN still collapses to a nearly-flat solution and still needs the soft penalty to track the finite-temperature reference. The dependence on perturbative guidance, and the difference in how the two methods make use of it, are therefore properties of the nonlinear loop-plus-thermal flow equation itself rather than artifacts of the Higgs–singlet mass mixing.
As a first end-to-end test of the pipeline we also solved the O ( 3 ) O(3) O ( 3 ) bounce equation on the preliminary 2D potential, obtaining S 3 ( T ) / T ≈ 3 × 10 2 S_3(T)/T \approx3 \times10^2 S 3 ( T ) / T ≈ 3 × 1 0 2 at T = 100 T = 100 T = 100 GeV against ≈ 6 × 10 2 \approx6 \times10^2 ≈ 6 × 1 0 2 for the perturbative finite-temperature reference; the gap is dominated by the un-converged 2D barrier and the number is quoted only as a demonstration (§ 6.3 ).
A quantitative comparison against an established nonperturbative result is not yet within reach for this model, on two counts. First, there is no truncation-free benchmark: the most complete existing nonperturbative treatment, a three-dimensional-EFT lattice study of the same Z 2 Z_2 Z 2 singlet model [12 ] , itself relies on a perturbative dimensional reduction to build the effective theory it then simulates. Second, and more fundamentally, our infrared potential is not yet independent of the perturbative guidance used in training: as discussed in § 6.4 , the mass sector is tied to an analytic thermal target, so comparing that sector against a perturbative or lattice benchmark would be partly circular, while the quartic/shape sector, though unconstrained, is not yet robust enough to compare quantitatively. Decoupling the flow result from the perturbative targets and then benchmarking it against Ref. [12 ] is left for future work.
Several directions follow naturally from this work. On the numerical side, bringing the two-dimensional PINN to the same level of convergence as the one-dimensional results, and using its analytic barrier and bounce action along general paths in field space, would turn this demonstration into a fully nonperturbative determination of the nucleation temperature and transition strength. Removing the PINN’s remaining dependence on hand-tuned consistency-loss weights – whether through a better-conditioned architecture, a more informative choice of collocation points, or a principled rather than hand-tuned weighting scheme – remains an open problem. On the physics side, extending the present Z 2 Z_2 Z 2 -symmetric, CP-conserving truncation to include the CP-violating interactions needed for a complete treatment of electroweak baryogenesis is a natural next step, as is applying the same mesh-free FRG approach to other multi-field BSM extensions where a fully nonperturbative effective potential is needed.
Code and data availability
The source code for the physics-informed neural network training, the grid-based Newton–Krylov relaxation solver, and the scripts that generate the figures in this paper are openly available at https://github.com/nyokozaki/frg_pinns_higgs_singlet_pub. The trained network weights, the loss-weight configurations, and the benchmark input parameters used for the results reported here are attached to the tagged v1.0 release of that repository.
Acknowledgments
This work is supported by a startup grant from Zhejiang University.
Implementation details
Table 3 collects the network, optimizer, and loss-function hyperparameters used for the finite-temperature runs of § 6 . Entries that differ between the two benchmark temperatures are written as T = 100 ∣ T = 200 T=100 \mid T=200 T = 100 ∣ T = 200 . The complete per-block configuration files for every run are included in the code release.
Table 3. Network, optimizer, and loss hyperparameters for the finite-temperature PINN runs of § 6 . “rel. weight” denotes the weight of a combination inside the average defining L s i g n \mathcal{L}_{\mathrm{sign}} L sign or L m a g \mathcal{L}_{\mathrm{mag}} L mag , before multiplication by the outer w s i g n w_{\mathrm{sign}} w sign , w m a g w_{\mathrm{mag}} w mag . Combinations not listed in the last block carry no penalty at finite temperature.
Two-loop renormalization group equations
The two-loop RGEs used to run the couplings between k I R k_{\mathrm{IR}} k IR and k U V k_{\mathrm{UV}} k UV (§ 4.1 ) and to build the perturbative targets of the soft consistency penalty (§ 5.2 ) were generated with PyR@TE3 [23 ] for the Z 2 Z_{2} Z 2 -symmetric model of Eq. (16) . Their one-loop parts agree with those quoted in the singlet-extension literature [46 ] , and the two-loop structure follows the general results of Refs. [47 , 48 , 49 ] . We use the convention
β ( X ) ≡ μ d X d μ = 1 ( 4 π ) 2 β ( 1 ) ( X ) + 1 ( 4 π ) 4 β ( 2 ) ( X ) , (57) \beta(X) \equiv\mu\frac{dX}{d\mu} = \frac{1}{(4\pi)^2}\beta^{(1)}(X) + \frac{1}{(4\pi)^4}\beta^{(2)}(X), \tag*{(57)} β ( X ) ≡ μ d μ d X = ( 4 π ) 2 1 β ( 1 ) ( X ) + ( 4 π ) 4 1 β ( 2 ) ( X ) , (57)
keep only the top Yukawa coupling y t y_t y t (so Y u = diag ( 0 , 0 , y t ) Y_u = \operatorname{diag}(0, 0, y_t) Y u = diag ( 0 , 0 , y t ) ), and write μ H 2 , μ S 2 \mu_H^2,\mu_S^2 μ H 2 , μ S 2 for the Higgs and singlet mass parameters and λ H , λ S , λ H S \lambda_H,\lambda_S,\lambda_{HS} λ H , λ S , λ H S for the quartics of Eq. (16) .
Gauge couplings.
β ( 1 ) ( g 1 ) = 41 6 g 1 3 \beta^{(1)}(g_1) = \frac{41}{6}g_1^3 β ( 1 ) ( g 1 ) = 6 41 g 1 3
β ( 2 ) ( g 1 ) = + 199 18 g 1 5 + 9 2 g 1 3 g 2 2 + 44 3 g 1 3 g 3 2 − 17 6 g 1 3 y t 2 \beta^{(2)}(g_1)=+\frac{199}{18}g_1^5+\frac{9}{2}g_1^3g_2^2+\frac{44}{3}g_1^3g_3^2-\frac{17}{6}g_1^3y_t^2 β ( 2 ) ( g 1 ) = + 18 199 g 1 5 + 2 9 g 1 3 g 2 2 + 3 44 g 1 3 g 3 2 − 6 17 g 1 3 y t 2
β ( 1 ) ( g 2 ) = − 19 6 g 2 3 \beta^{(1)}(g_2)=-\frac{19}{6}g_2^3 β ( 1 ) ( g 2 ) = − 6 19 g 2 3
β ( 2 ) ( g 2 ) = + 3 2 g 1 2 g 2 3 + 35 6 g 2 5 + 12 g 2 3 g 3 2 − 3 2 g 2 3 y t 2 \beta^{(2)}(g_2)=+\frac{3}{2}g_1^2g_2^3+\frac{35}{6}g_2^5+12g_2^3g_3^2-\frac{3}{2}g_2^3y_t^2 β ( 2 ) ( g 2 ) = + 2 3 g 1 2 g 2 3 + 6 35 g 2 5 + 12 g 2 3 g 3 2 − 2 3 g 2 3 y t 2
β ( 1 ) ( g 3 ) = − 7 g 3 3 \beta^{(1)}(g_3)=-7g_3^3 β ( 1 ) ( g 3 ) = − 7 g 3 3
β ( 2 ) ( g 3 ) = + 11 6 g 1 2 g 3 3 + 9 2 g 2 2 g 3 3 − 26 g 3 5 − 2 g 3 3 y t 2 \beta^{(2)}(g_3)=+\frac{11}{6}g_1^2g_3^3+\frac{9}{2}g_2^2g_3^3-26g_3^5-2g_3^3y_t^2 β ( 2 ) ( g 3 ) = + 6 11 g 1 2 g 3 3 + 2 9 g 2 2 g 3 3 − 26 g 3 5 − 2 g 3 3 y t 2
Top Yukawa coupling.
β ( 1 ) ( y t ) = + 9 2 y t 3 − 17 12 g 1 2 y t − 9 4 g 2 2 y t − 8 g 3 2 y t \beta^{(1)}(y_t)=+\frac{9}{2}y_t^3-\frac{17}{12}g_1^2y_t-\frac{9}{4}g_2^2y_t-8g_3^2y_t β ( 1 ) ( y t ) = + 2 9 y t 3 − 12 17 g 1 2 y t − 4 9 g 2 2 y t − 8 g 3 2 y t
β ( 2 ) ( y t ) = − 12 y t 5 − 12 λ H y t 3 + 6 λ H 2 y t + 1 4 λ H S 2 y t + 131 16 g 1 2 y t 3 + 225 16 g 2 2 y t 3 + 36 g 3 2 y t 3 + 1187 216 g 1 4 y t − 3 4 g 1 2 g 2 2 y t + 19 9 g 1 2 g 3 2 y t − 23 4 g 2 4 y t + g 2 2 g 3 2 y t − 108 g 3 4 y t \begin{aligned}
\beta^{(2)}(y_t)={}&-12y_t^5-12\lambda_Hy_t^3+6\lambda_H^2y_t+\frac{1}{4}\lambda_{HS}^2y_t+\frac{131}{16}g_1^2y_t^3+\frac{225}{16}g_2^2y_t^3+36g_3^2y_t^3+\frac{1187}{216}g_1^4y_t\\
&-\frac{3}{4}g_1^2g_2^2y_t+\frac{19}{9}g_1^2g_3^2y_t-\frac{23}{4}g_2^4y_t+g_2^2g_3^2y_t-108g_3^4y_t
\end{aligned} β ( 2 ) ( y t ) = − 12 y t 5 − 12 λ H y t 3 + 6 λ H 2 y t + 4 1 λ H S 2 y t + 16 131 g 1 2 y t 3 + 16 225 g 2 2 y t 3 + 36 g 3 2 y t 3 + 216 1187 g 1 4 y t − 4 3 g 1 2 g 2 2 y t + 9 19 g 1 2 g 3 2 y t − 4 23 g 2 4 y t + g 2 2 g 3 2 y t − 108 g 3 4 y t
Scalar quartic couplings.
β ( 1 ) ( λ H ) = + 24 λ H 2 + 1 2 λ H S 2 − 3 g 1 2 λ H − 9 g 2 2 λ H + 3 8 g 1 4 + 3 4 g 1 2 g 2 2 + 9 8 g 2 4 + 12 λ H y t 2 − 6 y t 4 \beta^{(1)}(\lambda_H)=+24\lambda_H^2+\frac{1}{2}\lambda_{HS}^2-3g_1^2\lambda_H-9g_2^2\lambda_H+\frac{3}{8}g_1^4+\frac{3}{4}g_1^2g_2^2+\frac{9}{8}g_2^4+12\lambda_Hy_t^2-6y_t^4 β ( 1 ) ( λ H ) = + 24 λ H 2 + 2 1 λ H S 2 − 3 g 1 2 λ H − 9 g 2 2 λ H + 8 3 g 1 4 + 4 3 g 1 2 g 2 2 + 8 9 g 2 4 + 12 λ H y t 2 − 6 y t 4
β ( 2 ) ( λ H ) = − 312 λ H 3 − 5 λ H S 2 λ H − 2 λ H S 3 + 36 g 1 2 λ H + 108 g 2 2 λ H + 629 24 g 1 4 λ H + 39 4 g 1 2 g 2 2 λ H − 73 8 g 2 4 λ H − 379 48 g 1 6 − 559 48 g 1 4 g 2 2 − 289 48 g 1 2 g 2 4 + 305 16 g 2 6 − 144 λ H 2 y t 2 + 85 6 g 1 2 λ H y t 2 + 45 2 g 2 2 λ H y t 2 + 80 g 3 2 λ H y t 2 − 19 4 g 1 4 y t 2 + 21 2 g 1 2 g 2 2 y t 2 − 9 4 g 2 4 y t 2 − 3 λ H y t 4 − 8 3 g 1 2 y t 4 − 32 g 3 2 y t 4 + 30 y t 6 \begin{aligned}
\beta^{(2)}(\lambda_H)={}&-312\lambda_H^3-5\lambda_{HS}^2\lambda_H-2\lambda_{HS}^3+36g_1^2\lambda_H+108g_2^2\lambda_H+\frac{629}{24}g_1^4\lambda_H+\frac{39}{4}g_1^2g_2^2\lambda_H\\
&-\frac{73}{8}g_2^4\lambda_H-\frac{379}{48}g_1^6-\frac{559}{48}g_1^4g_2^2-\frac{289}{48}g_1^2g_2^4+\frac{305}{16}g_2^6-144\lambda_H^2y_t^2+\frac{85}{6}g_1^2\lambda_Hy_t^2\\
&+\frac{45}{2}g_2^2\lambda_Hy_t^2+80g_3^2\lambda_Hy_t^2-\frac{19}{4}g_1^4y_t^2+\frac{21}{2}g_1^2g_2^2y_t^2-\frac{9}{4}g_2^4y_t^2-3\lambda_Hy_t^4-\frac{8}{3}g_1^2y_t^4\\
&-32g_3^2y_t^4+30y_t^6
\end{aligned} β ( 2 ) ( λ H ) = − 312 λ H 3 − 5 λ H S 2 λ H − 2 λ H S 3 + 36 g 1 2 λ H + 108 g 2 2 λ H + 24 629 g 1 4 λ H + 4 39 g 1 2 g 2 2 λ H − 8 73 g 2 4 λ H − 48 379 g 1 6 − 48 559 g 1 4 g 2 2 − 48 289 g 1 2 g 2 4 + 16 305 g 2 6 − 144 λ H 2 y t 2 + 6 85 g 1 2 λ H y t 2 + 2 45 g 2 2 λ H y t 2 + 80 g 3 2 λ H y t 2 − 4 19 g 1 4 y t 2 + 2 21 g 1 2 g 2 2 y t 2 − 4 9 g 2 4 y t 2 − 3 λ H y t 4 − 3 8 g 1 2 y t 4 − 32 g 3 2 y t 4 + 30 y t 6
β ( 1 ) ( λ S ) = + 18 λ S 2 + 2 λ H S 2 \beta^{(1)}(\lambda_S)=+18\lambda_S^2+2\lambda_{HS}^2 β ( 1 ) ( λ S ) = + 18 λ S 2 + 2 λ H S 2
β ( 2 ) ( λ S ) = − 204 λ S 3 − 20 λ H S 2 λ S − 8 λ H S 3 + 4 g 1 2 λ H S 2 + 12 g 2 2 λ H S 2 − 12 λ H S y t 2 \beta^{(2)}(\lambda_S)=-204\lambda_S^3-20\lambda_{HS}^2\lambda_S-8\lambda_{HS}^3+4g_1^2\lambda_{HS}^2+12g_2^2\lambda_{HS}^2-12\lambda_{HS}y_t^2 β ( 2 ) ( λ S ) = − 204 λ S 3 − 20 λ H S 2 λ S − 8 λ H S 3 + 4 g 1 2 λ H S 2 + 12 g 2 2 λ H S 2 − 12 λ H S y t 2
β ( 1 ) ( λ H S ) = + 12 λ H S λ H + 6 λ H S λ S + 4 λ H S 2 − 3 2 g 1 2 λ H S − 9 2 g 2 2 λ H S + 6 λ H S y t 2 \beta^{(1)}(\lambda_{HS})=+12\lambda_{HS}\lambda_H+6\lambda_{HS}\lambda_S+4\lambda_{HS}^2-\frac{3}{2}g_1^2\lambda_{HS}-\frac{9}{2}g_2^2\lambda_{HS}+6\lambda_{HS}y_t^2 β ( 1 ) ( λ H S ) = + 12 λ H S λ H + 6 λ H S λ S + 4 λ H S 2 − 2 3 g 1 2 λ H S − 2 9 g 2 2 λ H S + 6 λ H S y t 2
β ( 2 ) ( λ H S ) = − 72 λ H S 2 λ H − 36 λ H S 2 λ S − 60 λ H S λ H 2 − 30 λ H S λ S 2 − 21 2 λ H S 3 + 24 g 1 2 λ H S λ H + 72 g 2 2 λ H S λ H + g 1 2 λ H S 2 + 3 g 2 2 λ H S 2 + 557 48 g 1 4 λ H S + 15 8 g 1 2 g 2 2 λ H S − 145 16 g 2 4 λ H S − 72 λ H S λ H y t 2 − 12 λ H S y t 4 + 85 12 g 1 2 λ H S y t 2 + 45 4 g 2 2 λ H S y t 2 + 40 g 3 2 λ H S y t 2 − 27 2 λ H S y t 4 \begin{aligned}
\beta^{(2)}(\lambda_{HS})={}&-72\lambda_{HS}^2\lambda_H-36\lambda_{HS}^2\lambda_S-60\lambda_{HS}\lambda_H^2-30\lambda_{HS}\lambda_S^2-\frac{21}{2}\lambda_{HS}^3+24g_1^2\lambda_{HS}\lambda_H\\
&+72g_2^2\lambda_{HS}\lambda_H+g_1^2\lambda_{HS}^2+3g_2^2\lambda_{HS}^2+\frac{557}{48}g_1^4\lambda_{HS}+\frac{15}{8}g_1^2g_2^2\lambda_{HS}-\frac{145}{16}g_2^4\lambda_{HS}\\
&-72\lambda_{HS}\lambda_Hy_t^2-12\lambda_{HS}y_t^4+\frac{85}{12}g_1^2\lambda_{HS}y_t^2+\frac{45}{4}g_2^2\lambda_{HS}y_t^2+40g_3^2\lambda_{HS}y_t^2-\frac{27}{2}\lambda_{HS}y_t^4
\end{aligned} β ( 2 ) ( λ H S ) = − 72 λ H S 2 λ H − 36 λ H S 2 λ S − 60 λ H S λ H 2 − 30 λ H S λ S 2 − 2 21 λ H S 3 + 24 g 1 2 λ H S λ H + 72 g 2 2 λ H S λ H + g 1 2 λ H S 2 + 3 g 2 2 λ H S 2 + 48 557 g 1 4 λ H S + 8 15 g 1 2 g 2 2 λ H S − 16 145 g 2 4 λ H S − 72 λ H S λ H y t 2 − 12 λ H S y t 4 + 12 85 g 1 2 λ H S y t 2 + 4 45 g 2 2 λ H S y t 2 + 40 g 3 2 λ H S y t 2 − 2 27 λ H S y t 4
Scalar mass parameters.
β ( 1 ) ( μ H 2 ) = − 3 2 g 1 2 μ H 2 − 9 2 g 2 2 μ H 2 + 12 λ H μ H 2 + λ H S μ S 2 + 6 μ H 2 y t 2 \beta^{(1)}(\mu_H^2)=-\frac{3}{2}g_1^2\mu_H^2-\frac{9}{2}g_2^2\mu_H^2+12\lambda_H\mu_H^2+\lambda_{HS}\mu_S^2+6\mu_H^2y_t^2 β ( 1 ) ( μ H 2 ) = − 2 3 g 1 2 μ H 2 − 2 9 g 2 2 μ H 2 + 12 λ H μ H 2 + λ H S μ S 2 + 6 μ H 2 y t 2
β ( 2 ) ( μ H 2 ) = + 557 48 g 1 4 μ H 2 + 15 8 g 1 2 g 2 2 μ H 2 − 145 16 g 2 4 μ H 2 + 24 g 1 2 λ H μ H 2 + 72 g 2 2 λ H μ H 2 − 60 λ H 2 μ H 2 − 1 2 λ H S 2 μ H 2 − 2 λ H S λ S μ S 2 + 85 12 g 1 2 μ H 2 y t 2 + 45 4 g 2 2 μ H 2 y t 2 + 40 g 3 2 μ H 2 y t 2 − 72 λ H μ H 2 y t 2 − 27 2 μ H 2 y t 4 \begin{aligned}
\beta^{(2)}(\mu_H^2)={}&+\frac{557}{48}g_1^4\mu_H^2+\frac{15}{8}g_1^2g_2^2\mu_H^2-\frac{145}{16}g_2^4\mu_H^2+24g_1^2\lambda_H\mu_H^2+72g_2^2\lambda_H\mu_H^2-60\lambda_H^2\mu_H^2\\
&-\frac{1}{2}\lambda_{HS}^2\mu_H^2-2\lambda_{HS}\lambda_S\mu_S^2+\frac{85}{12}g_1^2\mu_H^2y_t^2+\frac{45}{4}g_2^2\mu_H^2y_t^2+40g_3^2\mu_H^2y_t^2-72\lambda_H\mu_H^2y_t^2\\
&-\frac{27}{2}\mu_H^2y_t^4
\end{aligned} β ( 2 ) ( μ H 2 ) = + 48 557 g 1 4 μ H 2 + 8 15 g 1 2 g 2 2 μ H 2 − 16 145 g 2 4 μ H 2 + 24 g 1 2 λ H μ H 2 + 72 g 2 2 λ H μ H 2 − 60 λ H 2 μ H 2 − 2 1 λ H S 2 μ H 2 − 2 λ H S λ S μ S 2 + 12 85 g 1 2 μ H 2 y t 2 + 4 45 g 2 2 μ H 2 y t 2 + 40 g 3 2 μ H 2 y t 2 − 72 λ H μ H 2 y t 2 − 2 27 μ H 2 y t 4
β ( 1 ) ( μ S 2 ) = + 4 λ H S μ H 2 + 6 λ S μ S 2 \beta^{(1)}(\mu_S^2)=+4\lambda_{HS}\mu_H^2+6\lambda_S\mu_S^2 β ( 1 ) ( μ S 2 ) = + 4 λ H S μ H 2 + 6 λ S μ S 2
β ( 2 ) ( μ S 2 ) = + 8 g 1 2 λ H S μ H 2 + 24 g 2 2 λ H S μ H 2 − 8 λ H S 2 μ H 2 − 30 λ S 2 μ S 2 − 2 λ H S 2 μ S 2 − 24 λ H S μ H 2 y t 2 \beta^{(2)}(\mu_S^2)=+8g_1^2\lambda_{HS}\mu_H^2+24g_2^2\lambda_{HS}\mu_H^2-8\lambda_{HS}^2\mu_H^2-30\lambda_S^2\mu_S^2-2\lambda_{HS}^2\mu_S^2-24\lambda_{HS}\mu_H^2y_t^2 β ( 2 ) ( μ S 2 ) = + 8 g 1 2 λ H S μ H 2 + 24 g 2 2 λ H S μ H 2 − 8 λ H S 2 μ H 2 − 30 λ S 2 μ S 2 − 2 λ H S 2 μ S 2 − 24 λ H S μ H 2 y t 2