betaFeedback

WorksPhysics and Astronomy

A Systematic Benchmark of Physics-Informed Neural Network Architectures for the Stiff Poisson-Nernst-Planck System: Adaptive LossWeighting and Multi-Scale Resolution

Abstract

The Poisson Nernst Planck PNP system constitutes a canonical stiff coupled PDE problem where the charge density prefactor produces extreme coefficient ratios and the electric double layer imposes sharp boundary layers. Physics informed neural networks PINNs are appealing here because they require no mesh and differentiate through the physics automatically. Spectral bias and multi task loss imbalance however have limited their accuracy on stiff PNP systems. We present the first systematic data free benchmark of eleven PINN configurations organised into four strategy groups on a physically parametrised one dimensional PNP model for a lithium symmetric cell implemented within NVIDIA PhysicsNeMo Sym and validated against a finite volume method FVM reference. Root mean square errors RMSE span across architectures. The balanced residual decay rate BRDR scheme matches Neural Tangent Kernel NTK performance for concentration fields while reducing mean wall clock time making it the preferable strategy under compute constraints. Loss landscape geometry corroborates the RMSE ranking. We release an open source PhysicsNeMo Sym implementation for reuse on stiff coupled PDE problems in computational mechanics.

Introduction

The Poisson–Nernst–Planck (PNP) system is a prototypical example of a stiff, nonlinearly coupled PDE problem that challenges classical and modern numerical solvers alike. The dimensionless form of the system, derived in § 3, admits a singular-perturbation structure controlled by the parameter ε∝λD/L\varepsilon\propto\lambda_D/L, where λD\lambda_D is the Debye screening length and LL is the domain length. For typical electrochemical parameters, ε≪1\varepsilon\ll1 (ε≈10−4\varepsilon\approx10^{-4}–10−510^{-5} for concentrated electrolytes), which forces sharp boundary layers at the electrode interfaces that classical finite element codes address through Slotboom variable transformations, Gummel-iteration decoupling, and locally refined meshes [50]. At the same time, the extreme charge-density prefactor F/ε0≈1.09×1016 C V−1 m−3F/\varepsilon_0 \approx1.09 \times10^{16}\ \mathrm{C}\,\mathrm{V}^{-1}\,\mathrm{m}^{-3} creates coefficient ratios that destabilize standard iterative solvers [2]. The physical application motivating this work is ion transport in a one-dimensional lithium symmetric cell with LiPF6_6 electrolyte, which provides a parametrically well-documented and experimentally validated instance of the PNP system [35, 40, 48], and the same equations govern ion channels [12], nanofluidic devices [8], and solid-state ionic transport [2].

Neural networks for solving PDEs were first proposed by Lagaris et al. [26]; the modern physics-informed neural network (PINN) framework of Raissi et al. [38] placed this on a rigorous automatic differentiation footing and has since become a leading paradigm in scientific machine learning [6, 11, 22]. PINNs are attractive for stiff coupled PDEs because they require no mesh, differentiate through physical laws automatically, and accommodate operator-learning extensions [43]. However, two fundamental difficulties limit their accuracy on PNP-class problems. The first is spectral bias: PINNs preferentially learn low-frequency components of the solution, making the stiff, higher-frequency Poisson equation difficult to resolve [25, 37, 51]. The second is multi-task loss imbalance: because the PNP system couples equations with vastly different characteristic scales, the individual loss components converge at different rates, and naive uniform weighting causes the optimizer to over-satisfy the smoother Nernst–Planck equations at the expense of the stiffer Poisson equation [42, 45, 46]. Rigorous error bounds for PINNs on elliptic and parabolic PDEs have been established by Shin et al. [39] and generalisation error estimates by De Ryck and Mishra [9], Mishra and Molinaro [32, 33]; however, none of these analyses address the regime of extreme coefficient ratios (ε2∼10−10\varepsilon^2 \sim10^{-10}) characteristic of the PNP system studied here.

Prior PINN work on PNP has not adequately resolved either difficulty at practically useful accuracy. Huang et al. [17] proposed data-free enriched PINNs for dynamic PNP systems and demonstrated accuracy improvements over vanilla PINNs, but did not conduct a systematic multi-architecture benchmark under battery-relevant parametrisation. No systematic, data-free, multi-architecture benchmark exists for the PNP system under battery-relevant parametrisation.

This paper closes that gap. We benchmark eleven PINN configurations, organised into four strategy groups, against a validated FVM reference on the one-dimensional PNP system of Subramanian et al. [40], implemented entirely within NVIDIA’s open-source PhysicsNeMo Sym framework [36]. The four groups are: adaptive loss weighting (NTK, BRDR, AdaHessian), which corrects gradient pathologies without changing the network architecture; spectral bias mitigation (Fourier features, PIKAN), which modifies the network’s input representation or basis functions to resolve high-frequency boundary layers; spatio-temporal decomposition (FBPINN, Decoupled, SPINN, Sym./antisym. transform), which reduces problem complexity by splitting the domain, the equations, or the solution variables; and physics enrichment (EPINN), which embeds stiff analytical features directly into the network basis. The specific contributions are as follows. We provide the first systematic, data-free benchmark of eleven PINN configurations spanning four methodological groups on a physically parametrised 1D PNP system, with errors averaged over ten independent training runs per configuration. We show that NTK adaptive loss weighting achieves RMSE as low as 6.6×10−46.6 \times10^{-4} (dimensionless), while BRDR weighting matches this within 9% at reduced computational cost. A loss landscape analysis provides geometric corroboration of the RMSE ranking; wall-time distributions characterise all eleven configurations on an NVIDIA H100 GPU. We also release an open-source, modular PhysicsNeMo Sym implementation for PNP problems and, more broadly, for stiff coupled PDE systems in computational electrochemistry.

Developments in Physics-Informed Neural Networks

Since Raissi et al. [38] introduced the modern PINN framework, spectral bias [37] and multi-task loss imbalance have been the main obstacles, and both are especially acute for the PNP system studied here. Spectral bias arises because standard multilayer perceptrons (MLPs) preferentially fit low-frequency, global trends, leaving high-frequency localised features such as the electric double layer (EDL) unresolved. Loss imbalance arises because stiff multiphysics systems impose loss components with disparate magnitudes and convergence rates, so that a uniform weighting neglects the stiffest equations.

Tancik et al. [41] show that Fourier feature mappings applied to the input layer enable learning of high-frequency components for low-dimensional problems; Wang et al. [44] extend this by analysing the eigenvalue structure of the resulting NTK and proposing modified Fourier networks. Wang et al. [42] identify gradient flow pathologies as the root cause of PINN training failure on multiphysics problems, and Wang et al. [45] develop NTK-based adaptive loss weighting, grounded in the infinite-width NTK theory of Jacot et al. [19], which equates convergence rates across loss components and substantially improves accuracy on multi-scale and nonlinear systems. Domain-decomposition approaches, originally proposed as conservative PINNs by Jagtap and Karniadakis [20] and extended to finite basis PINNs (FBPINNs) by Moseley et al. [34] with multilevel decomposition by Dolean et al. [10], address both spectral bias and the global-communication bottleneck between subdomains. Separable PINNs [5] reduce memory requirements by canonical polyadic (CP) tensor decomposition but sacrifice accuracy on problems where coupling between spatial dimensions is strong. Kolmogorov–Arnold networks (KANs) [28], implemented for PDE solving as PIKANs by Wang et al. [47], offer theoretical advantages over MLPs but exhibit increased inference cost due to basis function computation. McClenny and Braga-Neto [31] introduce per-collocation-point trainable weights (self-adaptive PINNs) as a further adaptive strategy. Chen et al. [4] propose the balanced residual decay rate (BRDR) method, demonstrating accuracy comparable to NTK weighting at lower computational cost. Second-order optimisation via AdaHessian [52] provides curvature information through Hutchinson’s estimator, offering an alternative convergence path. Our benchmark evaluates all of these strategies on the same PNP problem, providing the first head-to-head comparison on a stiff, nonlinearly coupled system with physical parametrisation.

Poisson–Nernst–Planck Model

Dimensional PNP system

Our base-case system is the one-dimensional PNP transport model of Subramaniam et al. [40], which describes the mean-field dynamics of a binary electrolyte (LiPF6_6) in a lithium symmetric cell. The model solves for the cation concentration CpC_p (Li+^+), anion concentration CnC_n (PF6−_6^-), and electric potential Φ\Phi. The dimensional governing equations, boundary conditions (BCs), and initial condition (IC) are collected in Table 1; physical parameters are given in Table 2. The cell geometry is illustrated in Figure 1.

Interior PDELeft BC (X=0X = 0)Right BCIC
∂2Φ∂X2=−Fε0εs(zpCp+znCn)\frac{\partial^2\Phi}{\partial X^2} = -\frac{F}{\varepsilon_0\varepsilon_s}(z_p C_p + z_n C_n)∂Φ∂X=0\frac{\partial\Phi}{\partial X} = 0Φ=0\Phi= 0Φ=0\Phi= 0
∂Cp∂T=Dp∂2Cp∂X2+zpDpFRT∂∂X(Cp∂Φ∂X)\frac{\partial C_p}{\partial T} = D_p\frac{\partial^2 C_p}{\partial X^2} + \frac{z_pD_pF}{RT}\frac{\partial}{\partial X}\left(C_p\frac{\partial\Phi}{\partial X}\right)−Dp∂Cp∂X−zpDpFRTCp∂Φ∂X=IappzpF-D_p\frac{\partial C_p}{\partial X} - \frac{z_pD_pF}{RT}C_p\frac{\partial\Phi}{\partial X} = \frac{I_{\mathrm{app}}}{z_pF}same as leftCp=C0C_p = C_0
∂Cn∂T=Dn∂2Cn∂X2+znDnFRT∂∂X(Cn∂Φ∂X)\frac{\partial C_n}{\partial T} = D_n\frac{\partial^2 C_n}{\partial X^2} + \frac{z_nD_nF}{RT}\frac{\partial}{\partial X}\left(C_n\frac{\partial\Phi}{\partial X}\right)−Dn∂Cn∂X−znDnFRTCn∂Φ∂X=0-D_n\frac{\partial C_n}{\partial X} - \frac{z_nD_nF}{RT}C_n\frac{\partial\Phi}{\partial X} = 0same as leftCn=C0C_n = C_0

Table 1. Dimensional PNP system. Interior PDE, boundary conditions (BCs), and initial condition (IC) for the cation concentration CpC_p, anion concentration CnC_n, and electric potential Φ\Phi. Both electrode interfaces impose identical flux BCs; the right BC for each field is identical in form to the left BC.

SymbolValueUnitsDescription
C0C_0500mol m−3^{-3}Initial electrolyte concentration
DpD_p4×10−104 \times10^{-10}m2^2 s−1^{-1}Diffusivity of Li+^+
DnD_n4×10−94 \times10^{-9}m2^2 s−1^{-1}Diffusivity of PF6−_6^-
zpz_p1—Cation charge number
znz_n−1—Anion charge number
IappI_{\mathrm{app}}7.72A m−2^{-2}Applied current density (chosen to give δ=0.3\delta= 0.3)
εs\varepsilon_s16.8—Relative permittivity of solvent
TT298.15KTemperature
LL7.5×10−47.5 \times10^{-4}mInter-electrode distance
TfT_f3 600sFinal simulation time
λD\lambda_D≈1.7×10−8\approx1.7 \times10^{-8}mDebye length (ε≈2.3×10−5\varepsilon\approx2.3 \times10^{-5})

Table 2. Physical parameters for LiPF6_6 in a lithium symmetric cell [40]. The diffusivity ratio Dn/Dp=10D_n/D_p = 10 and the small Debye-to-domain-length ratio ε≈2.3×10−5\varepsilon\approx2.3 \times10^{-5} are the primary sources of stiffness.

Schematic of the one-dimensional lithium symmetric cell

Figure 1. Schematic of the one-dimensional lithium symmetric cell. The computational domain (x∈[0,1]x \in[0,1]) spans the LiPF6_6 electrolyte between two identical lithium metal electrodes. Under the applied positive current density IappI_{\mathrm{app}}, the cation (Li+^+, cpc_p) migrates toward the cathode (x=1x = 1) driven by the electric field. The anion (PF6−_6^-, cnc_n) carries zero net flux at both electrode interfaces (no-flux boundary condition); it accumulates near x=0x = 0 as electromigration and diffusion balance inside the electric double layer. The electric potential φ\varphi is the third solved field. Thin bands at both interfaces represent the electric double layer of dimensionless thickness O(ε)≈2.3×10−5O(\varepsilon) \approx2.3 \times10^{-5}.

Nondimensionalisation and conditioning

Nondimensionalisation is essential for PINN training: dimensional variables span orders of magnitude that produce ill-conditioned loss functions and impede convergence [13, 30]. We adopt the scheme of Subramaniam et al. [40]:

cp=Cp/C0,cn=Cn/C0,φ=ΦF/(RT)(1)c_p = C_p/C_0,\quad c_n = C_n/C_0,\quad\varphi= \Phi F/(RT) \tag*{(1)}
x=X/L,t=TDp/L2(2)x = X/L,\quad t = \mathcal{T}D_p/L^2 \tag*{(2)}
ε=RTε0εszp2F2C0L2,δ=IappzpFC0Dp/L,ξ=DnDp(3)\varepsilon= \sqrt{\frac{RT\varepsilon_0\varepsilon_s}{z_p^2F^2C_0L^2}},\quad\delta= \frac{I_{\mathrm{app}}}{z_pFC_0D_p/L},\quad\xi= \frac{D_n}{D_p} \tag*{(3)}

where ε\varepsilon is the ratio of the Debye screening length λD\lambda_D to the domain length LL (the singular-perturbation parameter), δ\delta is the dimensionless applied current, and ξ=Dn/Dp=10\xi= D_n/D_p = 10 is the diffusivity ratio. For the parameters of Table 2 these evaluate to ε≈2.30×10−5\varepsilon\approx2.30 \times10^{-5} and δ=0.3\delta= 0.3; IappI_{\mathrm{app}} is set to 7.72 A m−27.72\ \mathrm{A}\ \mathrm{m}^{-2} (slightly below the 10 A m−210\ \mathrm{A}\ \mathrm{m}^{-2} of [40]) so that δ=0.3\delta= 0.3 serves as the canonical dimensionless current for all PINN and FVM computations in this benchmark. The dimensionless system is given in Table 3.

InteriorLeft BCRight BCIC
−ε2∂2φ∂x2=zpcp+zncn-\varepsilon^2 \dfrac{\partial^2\varphi}{\partial x^2} = z_pc_p + z_nc_n∂φ∂x=0\dfrac{\partial\varphi}{\partial x} = 0φ=0\varphi= 0φ=0\varphi= 0
∂cp∂t=∂2cp∂x2+zp∂∂x(cp∂φ∂x)\dfrac{\partial c_p}{\partial t} = \dfrac{\partial^2c_p}{\partial x^2} + z_p\dfrac{\partial}{\partial x}\left(c_p\dfrac{\partial\varphi}{\partial x}\right)−∂cp∂x−zpcp∂φ∂x=δ-\dfrac{\partial c_p}{\partial x} - z_pc_p\dfrac{\partial\varphi}{\partial x} = \delta−∂cp∂x−zpcp∂φ∂x=δ-\dfrac{\partial c_p}{\partial x} - z_pc_p\dfrac{\partial\varphi}{\partial x} = \deltacp=1c_p = 1
∂cn∂t=ξ[∂2cn∂x2+zn∂∂x(cn∂φ∂x)]\dfrac{\partial c_n}{\partial t} = \xi\left[\dfrac{\partial^2c_n}{\partial x^2} + z_n\dfrac{\partial}{\partial x}\left(c_n\dfrac{\partial\varphi}{\partial x}\right)\right]−∂cn∂x−zncn∂φ∂x=0-\dfrac{\partial c_n}{\partial x} - z_nc_n\dfrac{\partial\varphi}{\partial x} = 0−∂cn∂x−zncn∂φ∂x=0-\dfrac{\partial c_n}{\partial x} - z_nc_n\dfrac{\partial\varphi}{\partial x} = 0cn=1c_n = 1

Table 3. Dimensionless PNP system. Derived from the nondimensionalisation (1)–(3). The small parameter ε≈2.3×10−5\varepsilon\approx2.3 \times10^{-5} controls the sharpness of the electric double layer at x=0x = 0 and x=1x = 1.

The parameter ε≈2.30×10−5\varepsilon\approx2.30 \times10^{-5} is small, which means the Poisson equation −ε2φxx=zpcp+zncn-\varepsilon^2\varphi_{xx} = z_pc_p + z_nc_n has a coefficient ε2≈5.3×10−10\varepsilon^2 \approx5.3 \times10^{-10}: the electric potential must satisfy this equation with a coefficient ten orders of magnitude smaller than unity, creating the loss imbalance that is the central computational difficulty for all PINN configurations tested.

Finite volume reference solver

The reference solution is obtained by a method of lines (MOL) solver implemented in Python (pnp_fvm.py). The spatial domain x∈[0,1]x \in[0,1] is discretised on a uniform grid of NFVM=500N_{\mathrm{FVM}} = 500 cell-centred nodes with spacing h=1/NFVM=2×10−3h = 1/N_{\mathrm{FVM}} = 2 \times10^{-3}. At each time step the Poisson equation in Table 3 is solved algebraically as a (NFVM+1)×(NFVM+1)(N_{\mathrm{FVM}} + 1) \times(N_{\mathrm{FVM}} + 1) tridiagonal linear system (numpy.linalg.solve), with the left Neumann condition ∂xφ∣x=0=0\left.\partial_x\varphi\right|_{x=0} = 0 discretised by a first-order one-sided difference and the right Dirichlet condition φ(1,t)=0\varphi(1,t) = 0 imposed directly. Ion fluxes are approximated at cell faces using centered differences for concentration gradients and arithmetic-average face concentrations, consistent with the standard cell-centred finite-volume flux:

Np(j+12)=−cp(j+1)−cp(j)h−cˉp(j+12)φ(j+1)−φ(j)h,(4)N_p^{(j+\frac{1}{2})} = -\frac{c_p^{(j+1)} - c_p^{(j)}}{h} - \bar{c}_p^{(j+\frac{1}{2})}\frac{\varphi^{(j+1)} - \varphi^{(j)}}{h}, \tag*{(4)}

and analogously for NnN_n. The resulting stiff ODE system for the 2(NFVM+1)2(N_{\mathrm{FVM}}+1) concentration variables is integrated in time by scipy.integrate.solve_ivp with the Radau implicit Runge–Kutta solver (order 5, tolerances rtol =10−6= 10^{-6}, atol =10−8= 10^{-8}), which is well-suited to the stiff eigenvalue spectrum arising from the ε−2\varepsilon^{-2} coefficient in the Poisson equation. The boundary flux is set to δ=0.3\delta= 0.3 (Iapp≈7.72 A m−2I_{\mathrm{app}} \approx7.72\ \mathrm{A\,m^{-2}}), matching the value used in all PINN configurations. The solution is evaluated at 3 600 uniformly spaced times on [0,τf≈2.56][0,\tau_f \approx2.56], yielding a space-time dataset of 3 600×5013\,600 \times501 points for each of the three fields (cpc_p, cnc_n, φ\varphi) written to pnp.csv. The spatial discretisation error, estimated by Richardson extrapolation between N=250N = 250 and N=500N = 500 grids, is O(h2)≈4×10−6O(h^2) \approx4 \times10^{-6}, at least two orders of magnitude below the best PINN RMSE (6.6×10−46.6 \times10^{-4}), so the MOL solution is a reliable benchmark reference. A full derivation of the discrete equations, boundary conditions, and solver parameters is given in Supplementary Material Section 1. Existence and uniqueness of the continuous PNP solution for the parameter regime studied here follows from Jerome [21].

PINN Benchmark Design and Training Configurations

Physics-informed neural networks

Universal approximation theorems [7, 16] guarantee that a sufficiently expressive neural network can approximate any continuous function to arbitrary precision. PINNs exploit this by embedding physical laws directly into the loss function via automatic differentiation (AD), which removes the need for a mesh. Formally, the PINN problem (5)–(6) is:

θ∗=arg⁡min⁡θL(θ)(5)\theta^{*} = \arg\min_{\theta} \mathcal{L}(\theta) \tag*{(5)}
θn+1←θn−αn∇θL(θn)(6)\theta_{n+1} \leftarrow\theta_n - \alpha_n \nabla_{\theta}\mathcal{L}(\theta_n) \tag*{(6)}

where θ\theta denotes network parameters, αn\alpha_n is the learning rate at iteration nn, and Adam [24] provides adaptive per-parameter learning rates.

The composite loss for the PNP system is:

L(θ)=λPDELPDE+λBCLBC+λICLIC(7)\mathcal{L}(\theta) = \lambda_{\mathrm{PDE}}\mathcal{L}_{\mathrm{PDE}} + \lambda_{\mathrm{BC}}\mathcal{L}_{\mathrm{BC}} + \lambda_{\mathrm{IC}}\mathcal{L}_{\mathrm{IC}} \tag*{(7)}

where LPDE\mathcal{L}_{\mathrm{PDE}}, LBC\mathcal{L}_{\mathrm{BC}}, and LIC\mathcal{L}_{\mathrm{IC}} are the mean-squared PDE residual, boundary condition residual, and initial condition residual, respectively. The weights λi\lambda_i are either fixed at unity (vanilla) or adaptively determined at each training step by one of the two adaptive schemes described in Section 4.2.

PhysicsNeMo Sym framework

All implementations use NVIDIA’s open-source PhysicsNeMo Sym framework [36], formerly NVIDIA Modulus Sym, built on PyTorch. The framework has four features that are central to the benchmark design.

Computational graph and automatic differentiation. Computations are organised as a directed acyclic graph of Node objects, each encapsulating a network or a physics equation with named Key variables. The Graph class resolves inter-node dependencies and, in a single forward pass, computes all PDE residuals via backward-mode AD [1], without any symbolic manipulation.

Constraint system and Halton collocation. Typed constraint classes (InteriorConstraint, BoundaryConstraint, InitialCondition) encapsulate point sampling and loss computation. Points are sampled using quasirandom Halton sequences [14], which provide better space-filling coverage and O(N−1(log⁡N)d)O(N^{-1}(\log N)^d) quasi-Monte Carlo convergence compared to the O(N−1/2)O(N^{-1/2}) rate of independent uniform random sampling [3]. Interior and boundary points are resampled at every training iteration with a 10:1 interior-to-boundary ratio; the precise collocation counts for each configuration are reported in Table 4. We note that residual-based adaptive sampling strategies, which redistribute collocation points toward high-residual regions and have been shown to substantially improve accuracy for problems with sharp features [29, 49], were not employed in this benchmark to isolate the effect of architecture and loss-weighting choice. Given the boundary layer of dimensionless thickness O(ε)≈2.3×10−5O(\varepsilon) \approx2.3 \times10^{-5}, adaptive sampling is expected to yield additional accuracy gains and is identified as a priority for future work.

GroupConfigurationPrimary strength (key reference)NintN_{\mathrm{int}}NbndN_{\mathrm{bnd}}
BaselineVanilla PINNReference MLP; uniform loss weights [38]160001600
Adaptive WeightingNTK weightingFixes loss imbalance via NTK-based rescaling [45]160001600
Adaptive WeightingBRDR weightingFixes loss imbalance via balanced residual decay rate [4]160001600
Adaptive WeightingAdaHessianCurvature-aware optimizer; Hessian diagonal via Hutchinson [52]; halved collocation budget to offset the memory overhead of the Hessian-vector product8000800
Spectral Bias MitigationFourier featuresHigh-frequency input mapping lifts spectral bias [41]160001600
Spectral Bias MitigationPIKANKAN basis functions resolve multi-scale features [47]160001600
Spatio-Temporal DecompositionFBPINN (multilevel)Domain decomposition; localised subdomain networks [10]160001600
Spatio-Temporal DecompositionDecoupled PINNGummel-style equation splitting; reduces nonlinear coupling160001600
Spatio-Temporal DecompositionSPINNSeparable CP tensor decomposition; memory-efficient [5]160001600
Spatio-Temporal DecompositionSym./antisym. transf.Bazant variable decomposition reduces effective nonlinearity [2]160001600
EnrichmentEnriched PINN (EPINN)Physics-aware basis enrichment models stiff EDL features [17]160001600

Table 4. Summary of eleven PINN configurations organised into four strategy groups. Configurations sharing the same group exploit the same primary mechanism. Four entries constitute distinct network architectures (Vanilla, EPINN, FBPINN, PIKAN); the remaining seven modify the base MLP through optimizer, loss weighting, input encoding, or variable reformulation. NintN_{\mathrm{int}}: interior collocation points per step; NbndN_{\mathrm{bnd}}: boundary/initial condition (IC) points per step. All configurations share the base architecture (6 layers, 512 neurons, tanh) except where noted.

Adaptive loss weighting. The three weighting-group configurations address loss imbalance at different computational costs. Let L={L1,…,LN}\mathcal{L} = \{\mathcal{L}_1,\ldots,\mathcal{L}_N\} denote the set of scalar loss components (PDE, BC, IC residuals), θ∈RP\theta\in\mathbb{R}^{P} the network parameters, and λi\lambda_i the weight for component ii so that L=∑iλiLi\mathcal{L} = \sum_i \lambda_i \mathcal{L}_i.

NTK weighting [42, 45] is grounded in the infinite-width Neural Tangent Kernel theory of Jacot et al. [19]. In finite-width practice the per-constraint NTK trace is approximated by the squared gradient norm of the square-root loss, computed via a single backward pass:

Ki(NTK)=∥∂Li∂θ∥22=∑p=1P(∂Li∂θp)2(8)K_i^{(\mathrm{NTK})} = \left\lVert\frac{\partial\sqrt{\mathcal{L}_i}}{\partial\theta} \right\rVert_2^2 = \sum_{p=1}^{P} \left(\frac{\partial\sqrt{\mathcal{L}_i}}{\partial\theta_p}\right)^2 \tag*{(8)}

Weights are set so that each constraint drives parameter updates at the same effective rate:

Kˉ=1N∑j=1NKj(NTK),λi←KˉKi(NTK)(9)\bar{K} = \frac{1}{N}\sum_{j=1}^{N} K_j^{(\mathrm{NTK})}, \qquad\lambda_i \leftarrow\frac{\bar{K}}{K_i^{(\mathrm{NTK})}} \tag*{(9)}

A large Ki(NTK)K_i^{(\mathrm{NTK})} indicates that loss ii already drives large parameter updates; its weight is reduced to balance the slower-converging constraints. In this benchmark traces are recomputed every 10 gradient steps (see Algorithm 1).

BRDR weighting [4] replaces the gradient trace with cheaper scalar residual statistics. An exponential moving average (EMA) of the squared loss tracks each component’s running scale (the 4th-moment proxy), from which an inverse relative decay rate is formed:

m^i(n)=βcmi(n−1)+(1−βc)Li21−βcn,w^i(n)=Li(n)m^i(n)+ϵstab(10)\widehat{m}_i^{(n)} = \frac{\beta_c m_i^{(n-1)} + (1-\beta_c)\mathcal{L}_i^2}{1-\beta_c^n}, \qquad\widehat{w}_i^{(n)} = \frac{\mathcal{L}_i^{(n)}}{\sqrt{\widehat{m}_i^{(n)}+\epsilon_{\mathrm{stab}}}} \tag*{(10)}

A loss that decays slowly relative to its own running average receives a larger w^i\hat{w}_i. The weights are then normalised, EMA-smoothed, and rescaled to maintain a mean of unity across all constraints:

w~i(n)=w^i(n)wˉ(n)+ϵstab,λi(n)=βwλi(n−1)+(1−βw)w~i(n)1N∑j[βwλj(n−1)+(1−βw)w~j(n)]+ϵstab(11)\tilde{w}_i^{(n)} = \frac{\hat{w}_i^{(n)}}{\bar{w}^{(n)} + \epsilon_{\mathrm{stab}}}, \qquad\lambda_i^{(n)} = \frac{\beta_w \lambda_i^{(n-1)} + (1-\beta_w)\tilde{w}_i^{(n)}}{\frac{1}{N}\sum_j [\beta_w \lambda_j^{(n-1)} + (1-\beta_w)\tilde{w}_j^{(n)}] + \epsilon_{\mathrm{stab}}} \tag*{(11)}

with βc=βw=0.999\beta_c = \beta_w = 0.999 and ϵstab=10−14\epsilon_{\mathrm{stab}} = 10^{-14}. No backward pass beyond the standard gradient step is required.

AdaHessian [52] replaces the squared-gradient second moment of Adam with a diagonal Hessian curvature estimate. The diagonal of the Hessian ∇2L\nabla^2 \mathcal{L} is approximated via Hutchinson’s stochastic estimator [18] using a Rademacher random vector v∼U{±1}P\mathbf{v} \sim\mathcal{U}\{\pm1\}^P:

H~pp=vp(∇2L⋅v)p,p=1,…,P(12)\tilde{H}_{pp} = v_p \left(\nabla^2 \mathcal{L} \cdot\mathbf{v}\right)_p, \qquad p = 1,\ldots,P \tag*{(12)}

This is obtained without forming the full Hessian by computing a Hessian-vector product through a second-order backpropagation. The parameter update rule mirrors Adam but with H~pp2\tilde{H}_{pp}^2 in place of gp2g_p^2 for the second moment:

m^p(n)=β1mp(n−1)+(1−β1)gp(n)1−β1n(bias-corrected gradient EMA)(13)\hat{m}_p^{(n)} = \frac{\beta_1 m_p^{(n-1)} + (1-\beta_1)g_p^{(n)}}{1-\beta_1^n} \qquad\text{(bias-corrected gradient EMA)} \tag*{(13)}
v^p(n)=β2vp(n−1)+(1−β2)(H~pp(n))21−β2n(bias-corrected Hessian EMA)(14)\hat{v}_p^{(n)} = \frac{\beta_2 v_p^{(n-1)} + (1-\beta_2)(\tilde{H}_{pp}^{(n)})^2}{1-\beta_2^n} \qquad\text{(bias-corrected Hessian EMA)} \tag*{(14)}
θp(n+1)=θp(n)−αnm^p(n)v^p(n)+ϵ(parameter update)(15)\theta_p^{(n+1)} = \theta_p^{(n)} - \alpha_n \frac{\hat{m}_p^{(n)}}{\sqrt{\hat{v}_p^{(n)}}+\epsilon} \qquad\text{(parameter update)} \tag*{(15)}

where gp(n)=∂L/∂θpg_p^{(n)} = \partial\mathcal{L}/\partial\theta_p is the standard gradient. When curvature H~pp\tilde{H}_{pp} is large the effective step size is reduced, providing implicit pre-conditioning without storing the full Hessian matrix. Compared to NTK and BRDR, AdaHessian modifies the optimizer rather than the loss weights, and uses (β1,β2)=(0.9,0.999)(\beta_1,\beta_2) = (0.9, 0.999) as in Adam.

Gradient accumulation. To simulate larger effective batch sizes without proportional GPU memory cost, gradients are accumulated over Kacc=4K_{\mathrm{acc}} = 4 forward/backward passes before each optimizer step, effectively quadrupling the batch size.

The complete pipeline is summarised in Figure 2.

PhysicsNeMo Sym computational pipeline

Figure 2. PhysicsNeMo Sym computational pipeline. Halton collocation points are fed to the neural network architecture (left); automatic differentiation (AD) through the Graph computes PDE, boundary condition (BC), and initial condition (IC) residuals (centre-left); Neural Tangent Kernel (NTK) or balanced residual decay rate (BRDR) adaptive weighting rescales individual loss components before aggregation to total loss L\mathcal{L} (centre-right); gradients accumulate over Kacc=4K_{\mathrm{acc}} = 4 passes before the Adam or AdaHessian optimizer updates parameters θ\theta (right). The dashed arrow represents the backpropagation loop.

PINN architectures and training configurations

Table 4 organises the eleven configurations into four strategy groups plus a vanilla baseline. The grouping reflects the primary mechanism each configuration exploits, rather than the underlying network architecture: four configurations (Vanilla, EPINN, FBPINN, PIKAN) constitute distinct network architectures, while the remaining seven are modifications of the base MLP through optimizer choice (AdaHessian), loss weighting (NTK, BRDR), input-layer encoding (Fourier features), or variable/equation reformulation (Decoupled, SPINN, Sym./antisym. transform). A vanilla fully connected PINN (6 layers, 512 neurons per layer, tanh activation) serves as the baseline. All configurations except AdaHessian use the Adam optimizer [24] with initial learning rate α0=3×10−4\alpha_0 = 3 \times10^{-4} and exponential decay; weights are initialised by Kaiming uniform/normal [15]. Each model is trained for 100 000 epochs with gradient accumulation frequency Kacc=4K_{\mathrm{acc}} = 4; results are averaged over ten independent runs per configuration. PINN-FVM validation RMSE is computed every 1 000 epochs. The self-adaptive per-collocation-point weighting of McClenny and Braga-Neto [31] was not included among the eleven configurations due to incompatibility with the PhysicsNeMo Sym constraint API; this constitutes a limitation and a direction for future work.

Spectral bias mitigation: Fourier features and PIKAN

Fourier feature mapping [41] replaces the raw input x=(x,t)\mathbf{x}=(x,t) with a random sinusoidal embedding before the first MLP layer:

γ(x)=[cos⁡(2πBx),sin⁡(2πBx)]T,B∈Rm×2,Bjk∼N(0,σ2)(16)\gamma(\mathbf{x})=\left[\cos(2\pi B\mathbf{x}),\sin(2\pi B\mathbf{x})\right]^{\mathsf{T}},\qquad B\in\mathbb{R}^{m\times2},\qquad B_{jk}\sim\mathcal{N}(0,\sigma^2) \tag*{(16)}

By projecting inputs into a high-frequency feature space, the MLP backbone can represent the sharp EDL gradients that are suppressed by standard spectral bias [37, 44]. The implementation uses the PhysicsNeMo fourier architecture (6 layers, 512 neurons, sigmoid linear unit (SiLU) activation).

PIKAN [47] replaces the weight×\timesactivation structure of each MLP layer with learnable univariate spline maps. Each KAN layer computes:

xj(l+1)=∑iφij(l)(xi(l)),φij(x)=wbtanh⁡(x)+∑k=0G+pcijkBkp(x)(17)x_j^{(l+1)}=\sum_i\varphi_{ij}^{(l)}(x_i^{(l)}),\qquad\varphi_{ij}(x)=w_b\tanh(x)+\sum_{k=0}^{G+p}c_{ijk}B_k^p(x) \tag*{(17)}

where BkpB_k^p are B-spline basis functions of order p=3p=3 on a fixed uniform grid of G=32G=32 points over the normalised domain [−1,1]2[-1,1]^2, wbw_b is a learnable base-activation weight, and cijkc_{ijk} are learnable spline coefficients. The per-unit expressive power of the spline basis allows a shallower, narrower network (2 hidden layers, width 16) relative to the MLP baseline.

Spatio-temporal decomposition: FBPINN, SPINN, Decoupled, and symmetric/antisymmetric transformation

FBPINN [10] partitions the (x,t)∈[0,1]×[0,2.56](x,t)\in[0,1]\times[0,2.56] domain into a nℓ=3n_\ell=3-level hierarchy; level ll contains 2l−12^{l-1} overlapping subdomains (1+2+4=71+2+4=7 total). Each subdomain jj at level ll contributes through a small subnetwork ulju_{lj} weighted by a sigmoid window:

u^(x)=1nℓ∑l=1nℓ∑jwlj(x)ulj(x~lj)∑jwlj(x),(18)\hat{u}(\mathbf{x})=\frac{1}{n_\ell}\sum_{l=1}^{n_\ell}\frac{\sum_j w_{lj}(\mathbf{x})u_{lj}(\tilde{\mathbf{x}}_{lj})}{\sum_j w_{lj}(\mathbf{x})}, \tag*{(18)}
wlj(x)=σ(x−(clj−rlj)s)⋅σ((clj+rlj)−xs)(19)w_{lj}(\mathbf{x})=\sigma\left(\frac{\mathbf{x}-(\mathbf{c}_{lj}-\mathbf{r}_{lj})}{s}\right)\cdot\sigma\left(\frac{(\mathbf{c}_{lj}+\mathbf{r}_{lj})-\mathbf{x}}{s}\right) \tag*{(19)}

where x~lj=(x−clj)/rlj\tilde{x}_{lj} = (x - c_{lj})/r_{lj} is the subdomain-normalised input, cljc_{lj} and rljr_{lj} are the subdomain centre and half-width, ss is the sigmoid sharpness, and the overlap ratio is 2.7. Subnetworks use SiLU activation, 4 layers, and 32 neurons per layer.

SPINN [5] decomposes the solution via a CP tensor expansion of rank R=30R = 30:

u^(x,t)=∑r=1Rfr(x) gr(t)(20)\hat{u}(x,t) = \sum_{r=1}^{R} f_r(x)\,g_r(t) \tag*{(20)}

where frf_r and grg_r are scalar-valued MLP subnetworks (6 layers, 256 neurons, SiLU) applied to each coordinate independently. PDE derivatives require only 1D derivative computations through frf_r and grg_r, reducing memory at the cost of accuracy when spatial and temporal modes are strongly coupled, as is the case in the EDL regime.

The Decoupled PINN follows a Gummel-style sequential update: at each training step φ\varphi is optimised with fixed (cp,cn)(c_p,c_n), then (cp,cn)(c_p,c_n) are updated with the resulting φ\varphi. This reduces the effective nonlinearity of each sub-problem but introduces a splitting error that grows in the strongly coupled EDL.

For the symmetric/antisymmetric transformation [2], the Bazant-inspired variable change

ρ=12(cp−cn),(21)\rho= \frac{1}{2}(c_p - c_n), \tag*{(21)}
c=12(cp+cn)(22)c = \frac{1}{2}(c_p + c_n) \tag*{(22)}

decouples the two concentration equations at leading order, reducing the effective nonlinearity presented to the network. The complete derivation and the transformed PNP system are provided in the supplementary material.

Physics enrichment: EPINN

The Enriched PINN (EPINN) [17] applies two modifications to the baseline MLP. First, the activation function is changed from tanh to exponential linear unit (ELU) to improve gradient flow near the sharp EDL transition. Second, the loss is aggregated via homoscedastic uncertainty weighting, which introduces a trainable log-variance parameter σi\sigma_i for each constraint [23]:

Ltotal=∑i[Li2σi2+log⁡σi](23)\mathcal{L}_{\mathrm{total}} = \sum_i \left[\frac{\mathcal{L}_i}{2\sigma_i^2} + \log\sigma_i\right] \tag*{(23)}

The σi\sigma_i are co-optimized by Adam alongside the network weights. The log⁡σi\log\sigma_i regularisation term prevents all weights from collapsing to zero, providing principled loss balancing without requiring gradient traces (NTK) or residual statistics (BRDR).

Algorithm 1 describes the complete training workflow.

Results and Discussion

All eleven configurations converged within the 100 000-epoch budget. Dimensionless RMSE values span 10−210^{-2}–10−410^{-4} relative to the FVM reference. The complete error statistics, comprising mean RMSE and mean absolute error (MAE) averaged over ten independent runs per architecture, together with their standard deviations, are given in Table 5. Representative space-time solution surfaces and training loss curves for the best-performing NTK configuration are shown in Figure 3 and Figure 4.

Table 5: PINN-FVM dimensionless solution errors. Root-mean-square error (RMSE) and mean absolute error (MAE) averaged over ten independent training runs per configuration. Parenthetical values after RMSE entries are one standard deviation across the ten runs; MAE values are means only. Bold indicates the best result per metric per field. All values are dimensionless.

ConfigurationRMSE (std)MAE
Vanilla PINNcnc_n9.355×10−49.355 \times10^{-4} (0.8×10−40.8 \times10^{-4})5.019×10−45.019 \times10^{-4}
cpc_p8.511×10−48.511 \times10^{-4} (0.7×10−40.7 \times10^{-4})4.496×10−44.496 \times10^{-4}
φ\varphi1.585×10−31.585 \times10^{-3} (1.2×10−41.2 \times10^{-4})7.064×10−47.064 \times10^{-4}
AdaHessiancnc_n3.576×10−33.576 \times10^{-3} (0.5×10−30.5 \times10^{-3})2.567×10−32.567 \times10^{-3}
cpc_p3.454×10−33.454 \times10^{-3} (0.5×10−30.5 \times10^{-3})2.413×10−32.413 \times10^{-3}
φ\varphi6.729×10−36.729 \times10^{-3} (1.1×10−31.1 \times10^{-3})4.291×10−34.291 \times10^{-3}
BRDR weightingcnc_n7.270×10−47.270 \times10^{-4} (0.6×10−40.6 \times10^{-4})3.526×10−43.526 \times10^{-4}
cpc_p6.395×10−46.395 \times10^{-4} (0.5×10−40.5 \times10^{-4})3.198×10−43.198 \times10^{-4}
φ\varphi1.323×10−31.323 \times10^{-3} (1.0×10−41.0 \times10^{-4})4.484×10−44.484 \times10^{-4}
Decoupled PINNcnc_n5.651×10−35.651 \times10^{-3} (0.9×10−30.9 \times10^{-3})4.138×10−34.138 \times10^{-3}
cpc_p5.642×10−35.642 \times10^{-3} (0.8×10−30.8 \times10^{-3})4.122×10−34.122 \times10^{-3}
φ\varphi6.643×10−36.643 \times10^{-3} (1.3×10−31.3 \times10^{-3})5.000×10−35.000 \times10^{-3}
Enriched PINNcnc_n1.174×10−31.174 \times10^{-3} (0.9×10−40.9 \times10^{-4})6.930×10−46.930 \times10^{-4}
cpc_p1.129×10−31.129 \times10^{-3} (0.8×10−40.8 \times10^{-4})6.639×10−46.639 \times10^{-4}
φ\varphi1.717×10−31.717 \times10^{-3} (1.5×10−41.5 \times10^{-4})7.369×10−47.369 \times10^{-4}
FBPINN (multilevel)cnc_n4.642×10−34.642 \times10^{-3} (0.7×10−30.7 \times10^{-3})2.899×10−32.899 \times10^{-3}
cpc_p3.600×10−33.600 \times10^{-3} (0.6×10−30.6 \times10^{-3})2.230×10−32.230 \times10^{-3}
φ\varphi1.302×10−21.302 \times10^{-2} (2.1×10−32.1 \times10^{-3})7.450×10−37.450 \times10^{-3}
Fourier featurescnc_n4.758×10−34.758 \times10^{-3} (0.9×10−30.9 \times10^{-3})4.605×10−34.605 \times10^{-3}
cpc_p4.563×10−34.563 \times10^{-3} (0.8×10−30.8 \times10^{-3})4.399×10−34.399 \times10^{-3}
φ\varphi4.461×10−34.461 \times10^{-3} (0.7×10−30.7 \times10^{-3})4.060×10−34.060 \times10^{-3}
PIKANcnc_n1.419×10−31.419 \times10^{-3} (1.1×10−41.1 \times10^{-4})7.010×10−47.010 \times10^{-4}
cpc_p1.307×10−31.307 \times10^{-3} (1.0×10−41.0 \times10^{-4})6.258×10−46.258 \times10^{-4}
φ\varphi3.395×10−33.395 \times10^{-3} (0.5×10−30.5 \times10^{-3})1.541×10−31.541 \times10^{-3}
NTK weightingcnc_n6.613×10−46.613 \times10^{-4} (0.4×10−40.4 \times10^{-4})3.739×10−43.739 \times10^{-4}
cpc_p6.214×10−46.214 \times10^{-4} (0.3×10−40.3 \times10^{-4})3.225×10−43.225 \times10^{-4}
φ\varphi1.068×10−31.068 \times10^{-3} (0.1×10−30.1 \times10^{-3})4.180×10−44.180 \times10^{-4}
SPINNcnc_n2.448×10−22.448 \times10^{-2} (0.4×10−20.4 \times10^{-2})2.406×10−22.406 \times10^{-2}
cpc_p2.452×10−22.452 \times10^{-2} (0.4×10−20.4 \times10^{-2})2.410×10−22.410 \times10^{-2}
φ\varphi8.893×10−38.893 \times10^{-3} (1.5×10−31.5 \times10^{-3})3.755×10−33.755 \times10^{-3}
Sym./antisym. transformcnc_n8.992×10−48.992 \times10^{-4} (0.7×10−40.7 \times10^{-4})4.636×10−44.636 \times10^{-4}
cpc_p8.098×10−48.098 \times10^{-4} (0.6×10−40.6 \times10^{-4})4.216×10−44.216 \times10^{-4}
φ\varphi1.568×10−31.568 \times10^{-3} (1.1×10−41.1 \times10^{-4})6.229×10−46.229 \times10^{-4}

Table 5.

NTK-weighted PINN validation against the FVM reference

Figure 3. NTK-weighted PINN validation against the FVM reference. Space-time surface plots showing (left) the FVM reference solution, (centre) the PINN-predicted solution, and (right) the point-wise absolute difference for the dimensionless anion concentration cnc_n (top), cation concentration cpc_p (middle), and electric potential φ\varphi (bottom). Peak errors are localised near the electrode interfaces (x=0,1x = 0, 1) where the electric double layer imposes the sharpest gradients.

Average aggregated loss evolution and average loss component evolution

Figure 4. NTK-weighted PINN training loss. Total loss (left) and individual loss components (right), averaged over ten independent runs. Individual loss components (PDE: partial differential equation, BC: boundary condition, IC: initial condition) converge at comparable rates throughout training, consistent with the balanced weighting reported in § 4.2.

Adaptive loss weighting: NTK vs. BRDR

This subsection evaluates whether correcting loss imbalance via gradient-trace or residual-rate statistics is the dominant factor in PINN accuracy on the stiff PNP system, and quantifies the associated computational cost difference.

Both NTK and BRDR adaptive weighting substantially outperformed the vanilla PINN baseline (Table 5). The gain is consistent with loss imbalance, driven by ε2≈5.3×10−10\varepsilon^2 \approx5.3 \times10^{-10}, being the primary training bottleneck rather than architecture choice.

NTK weighting achieved the lowest RMSE across all three fields: (6.6±0.4)×10−4(6.6 \pm0.4) \times10^{-4} for cnc_n, (6.2±0.3)×10−4(6.2 \pm0.3) \times10^{-4} for cpc_p, and (1.1±0.1)×10−3(1.1 \pm0.1) \times10^{-3} for φ\varphi (mean ±\pm one standard deviation across ten runs). These values are 9%, 3%, and 19% below the corresponding BRDR RMSEs for cnc_n, cpc_p, and φ\varphi, respectively. However, BRDR achieved lower MAEs for both concentration fields, indicating that NTK suppresses large localised outliers more effectively (particularly near the electrode interfaces), whereas BRDR produces a more spatially uniform error distribution.

The NTK superiority in RMSE came at a mean additional cost of 3.2±0.43.2 \pm0.4 h wall-clock time per run (NVIDIA H100), because NTK requires computing the trace of the NTK matrix for each constraint at every step, whereas BRDR uses

Algorithm 1 PINN training loop for the 1D PNP system (PhysicsNeMo Sym)

Require: PNP parameters $(\varepsilon, \delta, \xi, z_p, z_n)$, architecture $\mathcal{A}$, weighting scheme $\mathcal{W}$, max epochs $N_{\mathrm{ep}} = 100\,000$, gradient accumulation frequency $K_{\mathrm{acc}} = 4$
Ensure: Trained parameters $\theta^*$, RMSE history vs. FVM reference

1: Initialize $\theta$ with Kaiming uniform/normal initialization
2: Set $\alpha_0 = 3 \times 10^{-4}$, exponential learning-rate decay (decay rate 0.92, steps 4000)
3: for epoch $n = 1$ to $N_{\mathrm{ep}}$ do
4:   for accumulation step $k = 1$ to $K_{\mathrm{acc}}$ do
5:     Sample $N_{\mathrm{int}}$ interior points $(x_i, t_i)$ via Halton sequence (see {xref}`Table 4`)
6:     Sample $N_{\mathrm{bnd}}$ points per boundary/IC constraint (see {xref}`Table 4`)
7:     Forward pass: $(\hat{c}_p, \hat{c}_n, \hat{\varphi}) = \mathrm{NN}_{\theta}(x, t)$
8:     Compute PDE/BC/IC residuals via automatic differentiation
9:     if $\mathcal{W} = \mathrm{NTK}$ and $n \bmod 10 = 0$ then
10:       Compute NTK traces $\{\operatorname{tr}(K_i)\}$ for each constraint
11:       Set $\lambda_i \leftarrow \bar{\lambda}/\operatorname{tr}(K_i)$, where $\bar{\lambda} = \frac{1}{N}\sum_j \operatorname{tr}(K_j)$
12:     else if $\mathcal{W} = \mathrm{BRDR}$ then
13:       Update $\lambda_i$ via balanced residual decay rate rule {citep}`[4]`
14:     else
15:       $\lambda_i \leftarrow 1$ (uniform weights)
16:     end if
17:     Compute $\mathcal{L} = \lambda_{\mathrm{PDE}}\mathcal{L}_{\mathrm{PDE}} + \lambda_{\mathrm{BC}}\mathcal{L}_{\mathrm{BC}} + \lambda_{\mathrm{IC}}\mathcal{L}_{\mathrm{IC}}$
18:     Accumulate: $g += \nabla_{\theta}\mathcal{L}/K_{\mathrm{acc}}$
19:   end for
20:   Clip gradient norm: $g \leftarrow g \cdot \min(1, 0.5/\lVert g\rVert)$
21:   Update: $\theta \leftarrow \theta - \alpha_n g$ (Adam, $\beta_1 = 0.9$, $\beta_2 = 0.999$; or AdaHessian)
22:   if $n \bmod 1000 = 0$ then
23:     Compute RMSE against FVM reference; log to file
24:   end if
25: end for
26: return $\theta^* \leftarrow \theta$

only scalar residual statistics. When accuracy is the priority, NTK is the method of choice; when compute is constrained, BRDR delivers near-identical accuracy at lower cost.

Frequency-scale resolution and spectral bias

This subsection examines how effectively each architecture mitigates spectral bias in the high-frequency Poisson equation, and whether spectral bias reduction translates to lower total RMSE within the fixed training budget.

The point-wise electric potential error fields for all architectures are shown in Figure 1 of the Supplementary Material. The vanilla, BRDR-weighted, and NTK-weighted configurations exhibit φ\varphi errors roughly one order of magnitude larger than their concentration errors, with errors concentrated near x=0x = 0 and x=1x = 1, which is the signature of spectral bias in the Poisson equation, whose ε2\varepsilon^2 prefactor makes it the highest-frequency component of the system.

In contrast, the AdaHessian, decoupled, enriched, Fourier-featurized, PIKAN, and SPINN architectures resolve both frequency scales to a similar order of magnitude, with errors distributed more uniformly across the domain. These architectures are therefore more effective at mitigating spectral bias, consistent with theory [37, 51]. However, none of these architectures achieved the lowest total RMSE within the fixed 100 000-epoch budget, suggesting a convergence-speed trade-off: spectral bias mitigation allows resolution of high-frequency features, but at the cost of slower convergence to the low-frequency background that adaptive-weighting configurations learn first. Training to a precision target rather than a fixed epoch count would better reveal the ceiling accuracy of spectral-bias-aware architectures.

Symmetric/antisymmetric variable transformation

This subsection quantifies the accuracy benefit of the Bazant variable transformation independently of any loss-weighting strategy, isolating the effect of algebraic reformulation on the effective problem nonlinearity.

The decomposition into symmetric (cc) and antisymmetric (ρ\rho) concentration variables (21)–(22) yielded consistent, modest RMSE improvements over the vanilla PINN: 4% for cnc_n, 5% for cpc_p, and 1% for ϕ\phi. The relatively small gains indicate that, while the transformation reduces effective nonlinearity, the dominant accuracy bottleneck remains the loss imbalance rather than the algebraic form of the field variables. Combining this transformation with NTK or BRDR weighting is a natural direction for further accuracy gains.

Wall-clock time and computational efficiency

This subsection characterises the computational cost of each configuration to assess the accuracy–efficiency trade-off and identify bottlenecks for future optimisation.

Wall-time distributions are shown in Figure 5. The observed minimum wall time of 13.1 ± 0.3 h per run reflects the overhead of generating FVM-comparison validation snapshots every 1 000 epochs; production deployments with less frequent checkpointing would be substantially faster.

Wall time distributions for the NTK configuration and across all architectures

Figure 5. Wall-clock time distributions. Box plots of wall time for the NTK configuration over ten runs (left) and across all architectures (right). The red line marks the median; box edges denote the first and third quartiles; whiskers extend to 1.5×1.5\times the interquartile range (IQR). All runs performed on one NVIDIA H100 GPU.

SPINN was the fastest architecture owing to its separable forward-mode AD formulation [5], yet it produced the highest RMSE (∼10−2\sim10^{-2}), demonstrating that speed alone cannot substitute for adequate loss conditioning in stiff PNP problems. The multilevel FBPINN was the slowest, a consequence of its loop-based subdomain iteration; vectorising the subdomain sweeps is expected to recover competitive training speeds.

Loss landscape geometry

This subsection tests whether loss basin sharpness provides a geometry-based predictor of RMSE rank that does not require FVM comparison.

Loss landscapes are computed by evaluating the training loss over a two-dimensional grid of parameter perturbations along orthogonal random directions in weight space [27]. As shown in Figure 6, most architectures converged to sharp, well-defined basins. The NTK-weighted PINN produced the sharpest and most symmetric basin, consistent with its lowest RMSE. The SPINN landscape was notably flat and irregular, consistent with its highest RMSE and suggesting that its training trajectory settled in a broad, poorly conditioned minimum. We quantify basin sharpness by the ratio of the maximum to minimum loss value over a fixed L∞L_\infty ball of radius 10−210^{-2} in the perturbation directions; NTK achieves the lowest sharpness ratio of 1.8, while SPINN achieves 47.3, providing a geometric predictor of generalisation quality that correlates monotonically with the RMSE ranking in Table 5.

Loss landscape geometry across all PINN architectures

Figure 6. Loss landscape geometry for all PINN configurations. Two-dimensional landscapes are computed by perturbing trained parameters along orthogonal random directions in weight space [27]. A sharper, more symmetric basin (NTK, BRDR, vanilla) indicates a well-conditioned minimum with superior generalisation. The SPINN landscape is flat and irregular (sharpness ratio 47.3), consistent with its highest observed RMSE (Table 5). The NTK configuration achieves the sharpest basin (sharpness ratio 1.8).

Collocation density sensitivity

This subsection examines whether the NTK benchmark errors at Nint=16,000N_{\mathrm{int}} = 16{,}000 are sensitive to the interior collocation budget, and whether the reported RMSE values are collocation-converged.

Three additional single-seed NTK runs were performed at Nint∈{2,000,6,000,10,000}N_{\mathrm{int}} \in\{2{,}000, 6{,}000, 10{,}000\}, maintaining the 10:1 interior-to-boundary ratio used throughout the benchmark. Figure 7 shows the aggregated training loss for all three runs.

Aggregated loss evolution for three collocation densities

Figure 7. NTK training loss for three collocation densities. Aggregated training loss (log10_{10} scale) over 100 000 epochs for the NTK-weighted configuration at interior collocation counts Nint∈{2,000,6,000,10,000}N_{\mathrm{int}} \in\{2{,}000, 6{,}000, 10{,}000\} (10:1 interior-to-boundary ratio throughout). All three runs converge to a comparable final training loss of approximately 10−310^{-3}; point-wise RMSE against the FVM reference decreases monotonically with increasing NintN_{\mathrm{int}}.

All three runs converge to a final training loss of approximately 10−310^{-3} within the 100 000-epoch budget, and the loss curves are qualitatively similar across the three NintN_{\mathrm{int}} values. Inspection of the validator surfaces suggests that point-wise RMSE against the FVM reference decreases monotonically as NintN_{\mathrm{int}} increases from 2 000 to 10 000, consistent with the expectation that denser collocation better resolves the stiff electric double layer near the electrode interfaces.

Conclusions

This benchmark establishes that adaptive loss weighting is the dominant factor governing PINN accuracy on the stiff PNP system, outweighing architecture choice, input-space encoding, and variable reformulation. NTK weighting achieves dimensionless RMSE as low as (6.6±0.4)×10−4(6.6 \pm0.4) \times10^{-4} (anion), (6.2±0.3)×10−4(6.2 \pm0.3) \times10^{-4} (cation), and (1.1±0.1)×10−3(1.1 \pm0.1) \times10^{-3} (electric potential), establishing the strongest data-free accuracy reported for a 1D PNP system under battery-relevant parametrisation [17]. The cheaper BRDR scheme matches NTK within 10% of RMSE for the concentration fields (24% for ϕ\phi) by computing only scalar residual statistics rather than full NTK matrix traces, reducing mean wall time by 3.2±0.43.2 \pm0.4 h per run on an NVIDIA H100; NTK is the right choice when accuracy is the priority, BRDR when compute is constrained. Spectral-bias-aware architectures (Fourier features, FBPINN, PIKAN) produce more spatially uniform error distributions but do not achieve the lowest total RMSE within the fixed 100 000-epoch budget, indicating a convergence-speed trade-off that a precision-target or curriculum training protocol could resolve. Loss basin sharpness correlates monotonically with the RMSE ranking across all eleven configurations, providing an efficient geometry-based diagnostic that does not require FVM comparison. These findings generalise beyond the specific electrochemical application: the stiffness structure (small singular-perturbation parameter, inter-equation loss imbalance) is shared by semiconductor drift-diffusion, reactive porous-media transport, and coupled thermo-mechanical problems, and the same adaptive-weighting remedies are expected to transfer.

Acknowledgements

C.G.T.F. acknowledges the support of the Natural Sciences and Engineering Research Council of Canada (NSERC) [RGPIN-2024-03989]. This research was enabled in part by support provided by SHARCNET and the Digital Research Alliance of Canada.

Declaration of Competing Interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

The full PhysicsNeMo Sym PINN implementation, including all configuration files and post-processing scripts, is publicly available at https://github.com/Feugmo-Group/physicsnemo-sym-pnp.

References

  1. [1]Atılım Güneş Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: a survey. Journal of Machine Learning Research, 18(153):1–43, 2018. URL https://jmlr.org/papers/v18/17-468.html.
  2. [2]Martin Z Bazant, Katsuyo Thornton, and Armand Ajdari. Diffuse-charge dynamics in electrochemical systems. Physical Review E, 70(2):021506, 2004. doi: 10.1103/PhysRevE.70.021506.DOI
  3. [3]Russel E Caflisch. Monte carlo and quasi-monte carlo methods. Acta Numerica, 7:1–49, 1998. doi: 10.1017/S0962492900002804.DOI
  4. [4]Wenqian Chen, Amanda A. Howard, and Panos Stinis. Self-adaptive weights based on balanced residual decay rate for physics-informed neural networks and deep operator networks. Journal of Computational Physics, 542:114226, 2025. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2025.114226. URL https://www.sciencedirect.com/science/article/pii/S0021999125005091.
  5. [5]Junwoo Cho, Seungtae Nam, Hyunmo Yang, Seok-Bae Yun, Youngjoon Hong, and Eunbyung Park. Separable physics-informed neural networks. In A. Oh, T. Naumann, A. Glober­son, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 23761–23788. Cur­ran Associates, Inc., 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/file/4af827e7d0b7bdae6097d44977e87534-Paper-Conference.pdf.arxiv.org/abs/2306.15969
  6. [6]Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics-informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92(3):88, 2022. doi: 10.1007/s10915-022-01939-z.DOI
  7. [7]George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals and Systems, 2(4):303–314, 1989. doi: 10.1007/BF02551274.DOI
  8. [8]Hirofumi Daiguji, Peidong Yang, and Arun Majumdar. Ion transport in nanofluidic channels. Nano Letters, 4(1):137–142, 2004. doi: 10.1021/nl0348185.DOI
  9. [9]Tim De Ryck and Siddhartha Mishra. Error analysis for physics-informed neural networks (PINNs) approximating Kolmogorov PDEs. Advances in Computational Mathematics, 48(6):79, 2022. doi: 10.1007/s10444-022-09985-9.DOI
  10. [10]Victorita Dolean, Alexander Heinlein, Siddhartha Mishra, and Ben Moseley. Multilevel domain decomposition-based architectures for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 429:117116, 2024. ISSN 0045-7825. doi: https://doi.org/10.1016/j.cma.2024.117116. URL https://www.sciencedirect.com/science/article/pii/S0045782524003724.
  11. [11]Weinan E and Bing Yu. The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems. Communications in Mathematics and Statistics, 6(1):1–12, 2018. doi: 10.1007/s40304-018-0127-z.DOI
  12. [12]Robert S Eisenberg. Computing the field in proteins and channels. Journal of Membrane Biology, 150(1):1–25, 1996. doi: 10.1007/0023299000026.pubmed.ncbi.nlm.nih.gov/8699474
  13. [13]Ehsan Haghighat, Maziar Raissi, Adrian Moure, Hector Gomez, and Ruben Juanes. A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics. Computer Methods in Applied Mechanics and Engineering, 379:113741, 2021. doi: 10.1016/j.cma.2021.113741.DOI
  14. [14]John H Halton. On the efficiency of certain quasi-random sequences of points in evaluating multi-dimensional integrals. Numerische Mathematik, 2(1):84–90, 1960. doi: 10.1007/BF01386213.DOI
  15. [15]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1026–1034, 2015. doi: 10.1109/ICCV.2015.123.DOI
  16. [16]Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural Networks, 2(5):359–366, 1989. doi: 10.1016/0893-6080(89)90020-8.DOI
  17. [17]Xujia Huang, Fajie Wang, Benrong Zhang, and Hanqing Liu. Enriched physics-informed neural networks for dynamic poisson-nernst-planck systems. Mathematics and Computers in Simulation, 237:231–246, 2025. ISSN 0378-4754. doi: https://doi.org/10.1016/j.matcom.2025.04.037. URL https://www.sciencedirect.com/science/article/pii/S0378475425001752.
  18. [18]M. F. Hutchinson. A stochastic estimator of the trace of the influence matrix for Laplacian smoothing splines. Communications in Statistics – Simulation and Computation, 18(3):1059–1076, 1989. doi: 10.1080/03610918908812806.DOI
  19. [19]Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In Advances in Neural Information Processing Systems, volume 31, pages 8571–8580, 2018. URL https://proceedings.neurips.cc/paper/2018/hash/5a4be1fa34e62bb8a6ec6b91d2462f5a-Abstract.html.arxiv.org/abs/1806.07572
  20. [20]Ameya D Jagtap and George Em Karniadakis. Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems. Computer Methods in Applied Mechanics and Engineering, 365:113028, 2020. doi: 10.1016/j.cma.2020.113028.DOI
  21. [21]Joseph W Jerome. Analysis of Charge Transport: A Mathematical Study of Semiconductor Devices. Springer, Berlin, Heidelberg, 1996. doi: 10.1007/978-3-642-79987-7.DOI
  22. [22]George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422–440, 2021. doi: 10.1038/s42254-021-00314-5.DOI
  23. [23]Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482–7491, 2018. doi: 10.1109/CVPR.2018.00781.
  24. [24]Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2014. URL https://arxiv.org/abs/1412.6980.
  25. [25]Aditi Krishnapriyan, Amir Gholami, Shandian Zhe, Robert Kirby, and Michael W Mahoney. Characterizing possible failure modes in physics-informed neural networks. In Advances in Neural Information Processing Systems, volume 34, pages 26548–26560, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/df438e5206f31600e6ae4af72f2725f1-Abstract.html.arxiv.org/abs/2109.01050
  26. [26]Isaac E Lagaris, Aristidis Likas, and Dimitrios I Fotiadis. Artificial neural networks for solving ordinary and partial differential equations. IEEE Transactions on Neural Networks, 9(5):987–1000, 1998. doi: 10.1109/72.712178.DOI
  27. [27]Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. In Advances in Neural Information Processing Systems, volume 31, pages 6389–6399, 2018. URL https://proceedings.neurips.cc/paper/2018/hash/a41b3bb3e6b050b6c9067c67f663b915-Abstract.html.arxiv.org/abs/1712.09913
  28. [28]Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y. Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks, 2025. URL https://arxiv.org/abs/2404.19756.
  29. [29]Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. DeepXDE: A deep learning library for solving differential equations. SIAM Review, 63(1):208–228, 2021. doi: 10.1137/19M1274067.DOI
  30. [30]Stefano Markidis. The old and the new: Can physics-informed deep-learning replace traditional linear solvers? Frontiers in Big Data, 4:669097, 2021. doi: 10.3389/fdata.2021.669097.DOI
  31. [31]Levi D McClenny and Ulisses M Braga-Neto. Self-adaptive physics-informed neural networks. Journal of Computational Physics, 474:111722, 2023. doi: 10.1016/j.jcp.2023.111722.
  32. [32]Siddhartha Mishra and Roberto Molinaro. Estimates on the generalization error of physics-informed neural networks for approximating a class of inverse problems for PDEs. IMA Journal of Numerical Analysis, 42(2):981–1022, 2022. doi: 10.1093/imanum/drab032.DOI
  33. [33]Siddhartha Mishra and Roberto Molinaro. Estimates on the generalization error of physics-informed neural networks for approximating PDEs. IMA Journal of Numerical Analysis, 43(1):1–43, 2023. doi: 10.1093/imanum/drac032.
  34. [34]Ben Moseley, Andrew Markham, and Tarje Nissen-Meyer. Finite basis physics-informed neural networks (FBPINNs): a scalable domain decomposition approach for solving differential equations. Advances in Computational Mathematics, 49(4):62, 2023. doi: 10.1007/s10444-023-10065-9.DOI
  35. [35]John Newman and Karen E Thomas-Alyea. Electrochemical Systems. John Wiley & Sons, Hoboken, NJ, 3rd edition, 2004.
  36. [36]NVIDIA Corporation. PhysicsNeMo Sym: An open-source framework for physics-informed machine learning. https://github.com/NVIDIA/physicsnemo, 2024. Accessed: 2025.
  37. [37]Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks, 2019. URL https://arxiv.org/abs/1806.08734.
  38. [38]M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2018.10.045. URL https://www.sciencedirect.com/science/article/pii/S0021999118307125.
  39. [39]Yeonjong Shin, Jérôme Darbon, and George Em Karniadakis. On the convergence of physics informed neural networks for linear second-order elliptic and parabolic type PDEs. Communications in Computational Physics, 28(5):2042–2074, 2020. doi: 10.4208/cicp.OA-2020-0193.DOI
  40. [40]Anudeep Subramaniam, Jixian Chen, Taekyu Jang, Nicholas R Geise, Ryan M Kasse, Michael F Toney, and Venkat R Subramanian. Analysis and simulation of one-dimensional transport models for lithium symmetric cells. Journal of The Electrochemical Society, 166(15):A3806, 2019. doi: 10.1149/2.0261915jes.DOI
  41. [41]Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7537–7547. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper_files/paper/2020/file/55053683268957697aa39fba6f231c68-Paper.pdf.arxiv.org/abs/2006.10739
  42. [42]Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021. doi: 10.1137/20M1318043.DOI
  43. [43]Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed deeponets. Science Advances, 7(40):eabi8605, 2021. doi: 10.1126/sciadv.abi8605. URL https://www.science.org/doi/abs/10.1126/sciadv.abi8605.DOI
  44. [44]Sifan Wang, Hanwen Wang, and Paris Perdikaris. On the eigenvector bias of fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 384:113938, 2021. doi: 10.1016/j.cma.2021.113938.DOI
  45. [45]Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2021.110768. URL https://www.sciencedirect.com/science/article/pii/S00219991100663X.
  46. [46]Sifan Wang, Shyam Sankaran, and Paris Perdikaris. Respecting causality for training physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 421:116813, 2024. doi: 10.1016/j.cma.2024.116813.DOI
  47. [47]Yizheng Wang, Jia Sun, Jinshuai Bai, Cosmin Anitescu, Mohammad Sadegh Eshaghi, Xiaoying Zhuang, Timon Rabczuk, and Yinghua Liu. Kolmogorov–arnold-informed neural network: A physics-informed deep learning framework for solving forward and inverse problems based on kolmogorov–arnold networks. Computer Methods in Applied Mechanics and Engineering, 433:117518, 2025. ISSN 0045-7825. doi: https://doi.org/10.1016/j.cma.2024.117518. URL https://www.sciencedirect.com/science/article/pii/S0045782524007722.
  48. [48]Kevin N Wood, Eric Kazyak, Alexander F Chadwick, Kuan-Hung Chen, Ji-Guang Zhang, Katsuyo Thornton, and Neil P Dasgupta. Dendrites and pits: Untangling the complex behavior of lithium metal anodes through operando video microscopy. ACS Cent Sci, 2(11):790–801, October 2016.pubmed.ncbi.nlm.nih.gov/27924307
  49. [49]Chenxi Wu, Min Zhu, Qinyang Tan, Yen Kartha, and Lu Lu. A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 403:115671, 2023. doi: 10.1016/j.cma.2022.115671.DOI
  50. [50]Dexuan Xie and Zhen Chao. A finite element iterative solver for a pnp ion channel model with neumann boundary condition and membrane surface charge. Journal of Computational Physics, 423:109915, 2020. ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2020.109915. URL https://www.sciencedirect.com/science/article/pii/S0021999120306896.
  51. [51]Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain, 2019. URL https://arxiv.org/abs/1807.01251.
  52. [52]Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Mahoney. Adahessian: An adaptive second order optimizer for machine learning. Proceedings of the AAAI Conference on Artificial Intelligence, 35(12):10665–10673, May 2021. doi: 10.1609/aaai.v35i12.17275. URL https://ojs.aaai.org/index.php/AAAI/article/view/17275.DOI

Paper details

vlnc:W_t2xqjmtq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents