betaFeedback

WorksPhysics and Astronomy

Renormalization-Inspired Effective Field Neural Networks for Scalable Modeling of Classical and Quantum Many-Body Systems

Abstract

We introduce Effective Field Neural Networks (EFNNs), a new architecture based on continued functions – mathematical tools used in renormalization to handle divergent perturbative series. Our key insight is that neural networks can implement these continued functions directly, providing a principled approach to many-body interactions. Testing on three systems (a classical 3-spin infinite- range model, a continuous classical Heisenberg spin system, and a quantum double exchange model), we find that EFNN outperforms standard deep networks, ResNet, and DenseNet. Most striking is EFNN's generalization: trained on 10×1010 \times 10 lattices, it accurately predicts behavior on systems up to 40×4040\times 40 with no additional training – and the accuracy improves with system size, with a computational time speed-up of 10310^{3} compared to ED for 40×4040\times 40 lattice. This demonstrates that EFNN captures the underlying physics rather than merely fitting data, making it valuable beyond many-body problems to any field where renormalization ideas apply.

The collective behaviors of many interacting particles—such as spins, molecules, and atoms—give rise to some of the most intriguing phenomena in condensed matter physics. However, theoretical and computational studies of many-body problems often confront the curse of dimensionality and one must devise strategies to circumvent it. Machine learning has emerged as a powerful tool for extracting effective features from high-dimensional datasets and tackling long-standing problems in condensed matter physics. Numerous machine learning techniques have been applied to a wide range of classical and quantum many-body systems, including classifying phases of matter with supervised [1, 2, 3] and unsupervised machine learning [4, 5, 6, 7, 8], representing many-body quantum states [9–13], accelerating Monte Carlo simulations [14–17], fitting high-dimensional potential energy surfaces [18–22] and searching for spin glass ground states [23].

Standard deep neural networks (DNNs) are ineffective in solving many-body problems unless augmented with physical knowledge. For example, an 8×88 \times8 classical two-dimensional Ising model requires highly complex DNN structures [24]. Consequently, extensive research has focused on encoding physics into neural network architectures to enhance their effectiveness in physical applications, including incorporating physical laws and constraints [25–40]. These physics-encoded neural networks have found applications in computing effective Hamiltonians [41–43], property prediction [36, 39, 41, 43–49], structure search [37, 38], etc. These methods, although often outperforming standard DNNs, tend to rely on intuitive and simplified imitations of physical principles, and still remain elusive in revealing the underlying many-body interactions.

In this work, we propose Effective Field Neural Networks (EFNNs), a new architecture inspired by field theory. The core motivation is to decompose many-body interactions into single quasi-particle representations governed by an emergent effective field. Unlike standard DNNs where adjacent layers rely only on previous layers, EFNNs incorporate initial features at every layer, forming a self-similar structure trained via recursive self-refining that captures many-body interactions through renormalization principles. Unlike approaches such as FermiNet [50], which directly parameterizes high-dimensional wave-functions without explicit separation of quasi-particles and fields, our method does not presuppose a fixed ansatz for the wavefunction. Instead, the quasi-particle and effective field representations are learned through a recursive self-refining process, mirroring the renormalization group (RG) framework in field theory, while leveraging the expressive power of deep neural networks. This recursive approach progressively refines the accuracy of both the quasi-particle and effective field descriptions.

We first illustrate the EFNN architecture by reformulating the classical 2D Ising model within the EFNN framework. The Hamiltonian of a classical 2D Ising model is defined as H(S)=−J∑⟨i,j⟩sisjH(S)=-J\sum_{\langle i,j\rangle}s_i s_j, where ⟨i,j⟩\langle i,j\rangle denotes a summation over pairs of nearest-neighbor sites, si=±1s_i=\pm1 represents the spin at site ii, JJ is the interaction strength between two spins, and SS is the collection of all spins. We define the effective field on spin sis_i as ϕi(S)=−J2×\phi_i(S)=-\frac{J}{2}\times (sum of nearest neighbors), and the quasi-particle as siϕi(S)s_i\phi_i(S), yielding total energy H(S)=∑isiϕi(S)H(S)=\sum_i s_i\phi_i(S). As illustrated schematically in Fig. 1(a), the interacting spins are first mapped to an effective field, then the effective field element-wise multiplied with the interacting spins is mapped to independent spins. Finally, a summation is performed over the independent spins, and the total energy is obtained. We reformulate this as a neural network structure (Fig. 1(b)), where S0=SS_0 = S represents the interacting spins. The effective field layer F1F_1 and the quasi-particle layer S1S_1 constitute a single field-particle (FP) layer. The evaluation of S1S_1 relies on both the effective layer F1F_1 and the interacting spin layer S0S_0, therefore a connection is required from S0S_0 to S1S_1, as well as from F1F_1 to S1S_1, distinguishing this architecture from a standard DNN. Finally, a summation is performed over S1S_1 to obtain the energy EE.

Image not extracted

Figure 1. Energy evaluation of a 2D Ising model reformulated as a neural network. (a) Calculation of the effective field by summing the interacting spins. Each interacting spin is multiplied by its corresponding effective field to obtain independent spin values, and the total energy is determined by summing up these independent spins. (b) Neural network representation of the energy evaluation process. ⊙\odot represents the element-wise multiplication.

Image not extracted

Figure 2. Performance of EFNNs on a classical 3-spin infinite range model and classical Heisenberg spin system. (a) Architecture of the EFNN. (b) Performance of EFNNs, DenseNets, ResNets, DNNs on 3-spin infinite range model. (c) Performance of EFNNs, DenseNets, ResNets, DNNs on classical Heisenberg spin system.

In the classical 2D Ising model, both the effective field and quasi-particles can be exactly calculated, which, however, becomes infeasible for more complicated many-body interactions. Following the spirit of renormalization from field theory, we can recursively refine the evaluations of the effective field and the quasi-particles. This recursive structure, when combined with the multiplicative interactions and self-similar connectivity pattern, implements a continued function—a mathematical tool from renormalization theory for handling divergent perturbative series (see Supplementary Material [51] for the complete derivation). As shown in Fig. 2(a), we can extend the FP layer number in Fig. 1(b) to obtain the deep Effective Field Neural Networks. The evaluation procedure is detailed as

Fi=fi−1(Si−1), Si=gi(S0)⊙Fi, E=q(Sn),F_i = f_{i-1}(S_{i-1}),\ S_i = g_i(S_0) \odot F_i,\ E = q(S_n),

where the function fi−1f_{i-1} maps the previous quasi-particle layer Si−1S_{i-1} (for i≥2i \ge2) or the initial layer S0S_0 (for i=1i = 1) to the effective field layer FiF_i. The input layer S0S_0 is first processed by gig_i, then element-wise multiplied by FiF_i to produce the quasi-particle layer SiS_i. Each FP layer consists of FiF_i and SiS_i. The final quasi-particle layer SnS_n generates the output energy EE through the function qq, which sums all elements of the last layer. The functions fif_i are nonlinear (specifically, we use hyperbolic tangent functions) to implement renormalization by ensuring finite values, while gig_i are nonlinear to enhance expressiveness. Linear functions may suffice only for simpler models like the classical 2D Ising model with nearest-neighbor interactions.

We present three case studies to demonstrate the capability of EFNNs to capture many-body interactions. The first case study focuses on the classical 3-spin infinite-range model in 1D, with Hamiltonian:

H(S)=−∑i<j<kJijksisjsk,(1)H(S) = - \sum_{i<j<k} J_{ijk}s_i s_j s_k, \tag*{(1)}

where the spins take binary values {0,1}\{0,1\}, JijkJ_{ijk} are constants, and the system comprises 15 lattice sites. We train EFNNs to evaluate H(S)H(S) in Eq. (1), and compare its performance with those of DenseNets, ResNets, and DNNs.

We generate 16000 training and 4000 test samples using Eq. (1). The nonlinear mappings fif_i and gjg_j use sequential linear layers with tanh activations. The network is trained with the Adam optimizer and a decaying learning rate. Performance is evaluated by the relative error, defined as the root mean square error (RMSE) on the test set divided by the absolute value of the mean of H(S)H(S). We vary the number of network layers from 1 to 3 and the number of neurons from 15 to 150. The performance of EFNNs, DenseNets, ResNets, and DNNs is presented in Fig. 2(b). The relative error decreases with more neurons for all networks, but EFNNs significantly outperform others as layers increase, with the relative error decreasing below 5×10−35 \times10^{-3} for EFNNs with 2 and 3 layers when the number of neurons ≥105\ge10^5. The only exception is the ResNet with 3 layers and 150 neurons. In most cases, DenseNets, ResNets, and DNNs perform worse than 2-layer EFNNs.

Image not extracted

Figure 3. Symmetrization: Each of the three channels of S0S_0 (S0,xS_{0,x}, S0,yS_{0,y}, S0,zS_{0,z}) undergoes convolution and element-wise multiplication within the same channel. The resulting products are summed to form the symmetrization layer Tj(S0)T_j(S_0). Subsequently, Tj(S0)T_j(S_0) is passed through a sequence of layers—convolution, tanh activation, and another convolution—to produce F1F_1 for j=0j=0 or gj(Tj(S0))g_j(T_j(S_0)) for j=1,…,nj=1,\ldots,n.

To test the capability of EFNNs beyond discrete systems, our second case study examines a classical Heisenberg spin system, using the same three-body interaction model in Eq. (1) with continuous spins, where sis_i takes continuous values in [−1,1][-1,1] rather than discrete values 0, 1. We use the RMSE of the total energy as the performance metric, with results shown in Fig. 2(c), since the average value of Eq. (1) is 0. Compared to the discrete case, the configuration space is now continuous, yet EFNNs with 2 and 3 layers still outperform all other networks. Fig. 2(c) also shows the general trend that the error decreases with increasing number of neurons. The most accurate result is achieved by the 3-layer EFNN with 150 neurons, where the error of the total energy reduces to 4×10−24 \times10^{-2}. These results demonstrate that EFNNs effectively capture many-body interactions through the renormalization procedure of their recursive continued-function structure; failing to incorporate this structure leads to inaccurate results. We show in the Supplementary Material [51] that DNNs completely lack the structure of a continued function, while DenseNets and ResNets misrepresent a continued function.

We proceed to our third, more challenging case study: evaluating the Monte Carlo energy of the quantum double exchange model [52–54] at finite temperature. The Hamiltonian of an N×NN \times N lattice reads

H(S)=−t∑⟨i,j⟩,α(c^iα†c^jα+h.c.)−J2∑i,α,βs⃗i⋅c^iα†σ⃗αβc^iβ.(2)H(S) = -t \sum_{\langle i,j\rangle,\alpha} \left(\hat{c}^{\dagger}_{i\alpha}\hat{c}_{j\alpha} + \mathrm{h.c.}\right) - \frac{J}{2}\sum_{i,\alpha,\beta}\vec{s}_i \cdot\hat{c}^{\dagger}_{i\alpha}\vec{\sigma}_{\alpha\beta}\hat{c}_{i\beta}. \tag*{(2)}

where ⟨i,j⟩\langle i,j\rangle denotes nearest neighbors, c^iα\hat{c}_{i\alpha} is the electron annihilation operator with spin α\alpha at site ii, σ⃗\vec{\sigma} are the Pauli matrices, and s⃗i∈S2\vec{s}_i \in\mathbb{S}^2 is a 3D classical unit vector representing the local spin at site ii, the classical spin interacts with the electrons through an on-site coupling, and SS is the collection of all classical spins s⃗i\vec{s}_i. We set N=10N=10, J=16tJ=16t, μ=−8.3t\mu=-8.3t, T=0.1tT=0.1t, t=1t=1.

At finite temperature TT, the Monte Carlo energy EMC(S)=−T∑nlog⁡(1+e−En(S)/T)E_{\mathrm{MC}}(S)=-T\sum_n\log(1+e^{-E_n(S)/T}) determines the acceptance rate in Monte Carlo simulations, where En(S)E_n(S) are the eigenvalues of H(S)H(S) with chemical potential in Eq. (2). Computing the exact value of EMC(S)E_{\mathrm{MC}}(S) requires diagonalization of Eq. (2) at each Monte Carlo update step, incurring O(N6)O(N^6) computational complexity per step. With millions of Monte Carlo steps required for convergence, this approach becomes computationally prohibitive. An intuitive solution is to approximate EMC(S)E_{\mathrm{MC}}(S) using a perturbative expansion, yielding an effective model with RKKY-type interactions [55–57]. This effective model with 2-body interactions has proven sufficiently accurate to replace exact diagonalization in Determinant Quantum Monte Carlo (DQMC) simulations [58]. The effective model including up to 4-body interactions takes the form

Eeff(S)∼J0+∑ijgijs⃗i⋅s⃗j+∑ijklgijkl(s⃗i⋅s⃗j)(s⃗k⋅s⃗l).(3)E_{\mathrm{eff}}(S) \sim J_0 + \sum_{ij} g_{ij}\vec{s}_i \cdot\vec{s}_j + \sum_{ijkl} g_{ijkl}\left(\vec{s}_i \cdot\vec{s}_j\right)\left(\vec{s}_k \cdot\vec{s}_l\right). \tag*{(3)}

Here, the dot products include all pairs of spins on the lattice, including self-interactions. The perturbative approach faces a fundamental dilemma: truncating at low orders fails to capture the full physics, while including higher-order terms leads to numerical divergence due to the factorial growth in the number of terms. EFNN addresses this challenge through renormalization.

The spin interactions in Eq. (3) are exclusively expressed as dot products to satisfy the O(3)O(3) symmetry of Eq. (2). Inspired by this symmetry, we incorporate symmetrization layers into the EFNN architecture, depicted in Fig. 3. The symmetrization is achieved by summing dot products of each channel of S0S_0 after convolution:

Tj(S0)=∑k=x,y,zconv⁡(S0,k)⊙conv⁡(S0,k),j=0,1,…,n.T_j(S_0) = \sum_{k=x,y,z} \operatorname{conv}(S_{0,k}) \odot\operatorname{conv}(S_{0,k}), \quad j=0,1,\ldots,n.

Here, each S0,k∈RN×NS_{0,k} \in\mathbb{R}^{N \times N} is convolved into RC×N×N\mathbb{R}^{C \times N \times N}. Supplementary Material [51] details the convolution process and proves that TjT_j remains invariant under O(3)O(3) transformations. As shown in Fig. 4(a), in the first FP layer, F1F_1 and S1S_1 are initialized as

F1=f0(T0(S0)),S1=g1(T1(S0))⊙F1.F_1 = f_0(T_0(S_0)), \quad S_1 = g_1(T_1(S_0)) \odot F_1.
Image not extracted

Figure 4. Relative error of EFNNs. (a) Architecture of EFNN incorporating symmetrization layers TjT_j, j=0,1,…,nj = 0, 1, \ldots, n. (b) Performance of EFNNs, DenseNets, ResNets, DNNs and effective model on a test set with lattice size N=10N = 10, for channel numbers C=1,3,5,7,9C = 1, 3, 5, 7, 9. (c) Generalization performance of networks trained on 10×1010 \times10 lattices (C=3C = 3) when applied to larger systems.

For subsequent layers j=2,…,nj = 2, \ldots, n, the effective field and quasi-particle layers are

Fj=fj−1(Sj−1),Sj=gj(Tj(S0))⊙Fj.F_j = f_{j-1}(S_{j-1}), \quad S_j = g_j(T_j(S_0)) \odot F_j.

The mappings fjf_j and gkg_k consist of a convolutional layer, a tanh activation, and another convolutional layer. For the mapping q(Sn)=Eq(S_n) = E, the final quasi-particle layer SnS_n is convolved into a single-channel matrix, and an element-wise summation is performed to produce the scalar energy prediction EE, see Supplementary Material [51]. The layer dimensions are as follows: input layer S0∈R3×N×NS_0 \in\mathbb{R}^{3 \times N \times N}, intermediate layers FjF_j, Sj∈RC×N×NS_j \in\mathbb{R}^{C \times N \times N}, for j=1,…,nj = 1, \ldots, n, and output E=q(Sn)∈RE = q(S_n) \in\mathbb{R}. For comparison, we also incorporate symmetrization layers and final summation qq into DenseNet, ResNet, and DNN architectures, detailed in Supplementary Material [51].

We generate a total of 2×1062 \times10^6 data points and split them into 80% for training and 20% for testing. These data are obtained by exact diagonalization of the Hamiltonian in Eq. (2) with chemical potential. For the effective model in Eq. (3), we perform linear regression to obtain the coefficients. While the full effective model would require 12758826 parameters, we reduce this to 213201 by restricting interactions within each 5×55 \times5 tile. The EFNNs, DenseNets, ResNets, and DNNs with symmetrization are trained using networks with C=1,3,5,7,9C = 1, 3, 5, 7, 9 channels and n=1,2,3n = 1, 2, 3 layers, with parameter counts ranging from 200 to 30000.

with C=1,3,5,7,9C = 1, 3, 5, 7, 9 channels and n=1,2,3n = 1, 2, 3 layers, with parameter counts ranging from 200 to 30000. As shown in Fig. 4(b), the effective model achieves a relative error of approximately 2×10−22 \times10^{-2}, whereas the neural networks achieve relative errors on the order of 10−310^{-3}—approximately six times better using significantly fewer parameters. Among the neural networks, EFNNs with 1, 2, and 3 layers consistently outperform all other architectures, with some overlap observed between 1-layer EFNNs and 3-layer DNNs, as well as 1-, 2-, and 3-layer DenseNets at C=1,3C = 1, 3. For EFNNs, the error systematically decreases as CC increases and as the number of layers increases from 1 to 3. This superior performance with fewer parameters shows that the EFNNs possess inherent renormalization capabilities beyond simple parameter fitting.

We next test the generalization capability of these neural networks trained on 10×1010 \times10 lattices by evaluating their performance on systems up to 40×4040 \times40. We select networks with C=3C = 3 channels as a representative configuration and evaluate EFNN, DenseNet, ResNet, and DNN architectures across this range of lattice sizes. The convolution operations with padding under periodic boundary conditions (PBC) effectively capture interactions between localized and neighboring spins. As shown in Fig. 4(c), all neural network architectures show improved performance on larger lattices, with relative errors decreasing as the system size increases from 10×1010 \times10 to 40×4040 \times40. The EFNNs exhibit the best extrapolation capability and renormalization ability: the 3-layer EFNN achieves the lowest errors across all lattice sizes, reaching a remarkable relative error of approximately 4×10−44 \times10^{-4} at N=40N = 40. The 2-layer EFNN mostly outperforms the 1-layer EFNN, while among non-EFNN architectures, only the 3-layer DNN has comparable performance to the 1-layer EFNN. All other configurations consistently produce higher errors, confirming EFNN’s superior ability to capture scale-invariant features. Notably, as demonstrated in Supplementary Material [51], for N=40N = 40, EFNN achieves 10310^3 speed-up compared to ED.

We give some discussions and conclude this paper. In all three case studies—the classical 3-spin infinite-range model, continuous classical Heisenberg system, and quantum double exchange model—EFNNs consistently outperform DenseNets, ResNets, DNNs, and the effective model, achieving the lowest errors. Particularly striking is EFNN’s ability to generalize: trained on 10×1010 \times10 lattices, it accurately predicted energies for systems as large as 40×4040 \times40, while DenseNets, ResNets, and standard DNNs struggle with this extrapolation task. This superior performance isn’t accidental—it stems from EFNN’s mathematical foundation in continued functions. While conventional architectures like DenseNet and ResNet were designed through trial and error with various architectural innovations, EFNN’s structure directly implements continued functions, a mathematical framework that naturally captures many-body interactions up to infinite order. More importantly, for a perturbative expansion like Eq. (3), EFNN selects the most relevant terms via training, and expresses these terms in nonlinear continued function structure. Fig. 5 compares these structures, demonstrating that only EFNN (Fig. 5(a)) possesses the correct continued function structure. The Supplementary Material [51] provides a detailed mathematical explanation of why EFNN implements continued functions while ResNet and DenseNet do not. A special case of continued functions is the Padé approximant, which takes the form of a continued fraction and is ubiquitously used throughout physics and numerical methods for its rapid convergence, strong extrapolation capability, and compact representation [59]. Continued functions generalize continued fractions by replacing the inversion operation with generic nonlinear functions. Here, we use tanh activation functions instead of inverses, which makes EFNN trainable through standard backpropagation—something impossible with traditional Padé approximants due to gradient singularities. This simple substitution preserves the mathematical power of continued functions while enabling practical neural network training. Notably, EFNN actually implements self-similar approximation [60], a powerful mathematical framework that has been successfully applied across diverse fields including high energy physics [61–65], condensed matter physics [66, 67], quantum mechanics [60, 64, 68–72], quantum field theory [60, 73–75], and applied mathematics [66, 76–85]. Previously, applying self-similar approximation required tedious manual calculations to obtain order by order matching; EFNN changes this by automating the process through neural network optimization. EFNN’s self-similar structure also naturally comes from renormalization group equation, as detailed in the Supplementary Material [51].

Image not extracted

Figure 5. Comparison of neural network structures. (a) In an EFNN, the initial layer S0S_0 is recursively integrated into every quasi-particle layer. After a mapping, S0S_0 is multiplied with subsequent layers, forming a continued-function representation (see the equation above the figure). (b) A typical architecture of ResNet, only one skip connection starts from S0S_0 and summation is used, this is not a correct continued-function structure. (c) Removing the S0S_0 connections transforms an EFNN into a standard DNN, with no continued-function structure. (d) A typical architecture of DenseNet, CC is concatenation operation, this is not a correct continued-function structure.

In conclusion, we have developed EFNN, a neural network architecture inspired by field theory that captures high-order interactions in both classical and quantum models. The key to EFNN’s success lies in how it recursively refines effective fields and constructs quasi-particles to encode complex many-body effects. This partitioning into field and quasi-particle layers provides clear physical interpretation while maintaining computational efficiency. As we showed, the connections from the true particle layer S0S_0 to the quasi-particle layers create a continued function structure with renormalization properties similar to Padé approximants. What EFNN actually implements is self-similar approximation—a powerful mathematical technique that has been widely used in physics but previously required manual calculations for each problem. Once properly trained, this architecture learns effective representations that accurately predict energies even for systems with complicated many-body interactions. Our tests on three diverse systems—the classical 3-spin finite-range model, continuous classical Heisenberg spin system, and quantum double exchange model—demonstrate that EFNN consistently outperforms DenseNet, ResNet, and standard DNNs. Most remarkably, EFNNs trained on small systems improve accuracy on larger ones, opening possibilities for training on expensive DFT data from small lattices then inferring on much larger systems. Looking ahead, we plan to develop theoretical error bounds and apply EFNN to other challenging problems in physics.

DATA AVAILABILITY

The data that support the findings of this study are openly available in Zenodo at Ref. [86].

**Acknowledgments—**This work is supported by the Hong Kong Research Grants Council (16306220) and National Key R&D Program of China (2021YFA1401500). Y.Z. is supported by the Max Planck Partner lab on quantum materials from Max Planck Institute Chemical Physics of Solids.

References

Supplemental Information of Effective Field Neural Networks

supplementary_v2.pdf

In this supplementary information, we present the complete workflow of EFNN with symmetrization for the quantum double exchange model in Section I. We introduce the basics of Padé approximation, continued fractions, and continued functions in Section II, and demonstrate that EFNN implements a continued function, while ResNet, DenseNet and DNN are not continued functions. In Section III, we derive the structure of EFNN from the renormalization group (RG) equation. In Section IV, we show that the convolution layer ensures the system’s invariance under the point group O(3). In Section V, we present the adapted architectures of ResNet, DenseNet, and DNN for energy fitting of the quantum double exchange model. In Section VI, we break down the details of DQMC, and demonstrate how EFNN solves the computational bottleneck problem in DQMC for quantum double exchange model. We report the speed-up of EFNN against ED. In Section VII, we report the distributions of EFNN’s predictions, and the ground truth data, for quantum double exchange model, demonstrating that the accuracy of EFNN.

Computational workflow of EFNNs with symmetrization

Image not extracted

Figure 1 (PDF p. 9). Computational workflow of EFNNs with symmetrization

In this section, we present the complete workflow of EFNN with symmetrization for the quantum double exchange model. Fig. 1(a) illustrates the symmetrization process: each of the three channels of S0S_0 (S0,xS_{0,x}, S0,yS_{0,y}, S0,zS_{0,z}) undergoes convolution and element-wise multiplication within the same channel. The resulting products are summed to form the symmetrization layer Tj(S0)T_j(S_0). Subsequently, Tj(S0)T_j(S_0) is passed through a sequence of layers—convolution, tanh activation, and convolution—to produce F1F_1 for j=0j = 0 or gj(Tj(S0))g_j(T_j(S_0)) for j=1,…,nj = 1,\ldots,n. Fig. 1(a) is Fig. 3 in the main text. Fig. 1(b) shows the generation of quasi-particle layers: the transformed symmetrization layer gj(Tj(S0))g_j(T_j(S_0)) is element-wise multiplied with the effective field layer FjF_j, yielding the quasi-particle layer SjS_j for j=1,…,nj = 1,\ldots,n. Fig. 1(c) depicts the transformation to effective field layers: each quasi-particle layer Sk−1S_{k-1} is processed through convolution, tanh activation, and convolution to generate the effective field layer FkF_k for k=2,…,nk = 2,\ldots,n. Fig. 1(d) illustrates the energy evaluation: the final quasi-particle layer SnS_n is transformed into a single-channel matrix, and an element-wise summation of this matrix yields the scalar energy prediction EE. These visualizations are generated after training the model with C=5C = 5 channels and n=2n = 2 layers, using the 576th sample from the test set as input S0S_0.

Mathematical details of DenseNet, ResNet, DNN

In this section, we describe the mathematical structures of EFNN, ResNet, DenseNet, and DNN.

We denote the input spins as S0∈RNS_0 \in\mathbb{R}^{N}.

Image not extracted

Figure 2 (PDF p. 10). Comparison of structures of (a) EFNN, (b) ResNet, (c) DNN, and (d) DenseNet

Why EFNN is a continued function

In Fig. 2(a), the functions gkg_k, k=1,2,3k = 1, 2, 3, fjf_j, j=0,1,2j = 0, 1, 2, all have the forms of a composition of a linear mapping ∘\circ tanh⁡\tanh ∘\circ linear mapping. We denote

gk(z)=βktanh⁡(γkz+δk)+φk,k=1,2,3,(4)g_k(z) = \beta_k \tanh(\gamma_k z + \delta_k) + \varphi_k,\quad k = 1, 2, 3, \tag*{(4)}
fj(z)=dj+cjtanh⁡(Ajz+bj),j=0,1,2.(5)f_j(z) = d_j + c_j \tanh(A_j z + b_j),\quad j = 0, 1, 2. \tag*{(5)}

The quasi-particle layers in EFNN form a nested structure that implements a continued function. For instance, the third quasi-particle layer is expressed as:

S3=g3(S0)⊙f2{g2(S0)⊙f1[g1(S0)⊙f0(S0)]},(6)S_3 = g_3\left(S_0\right) \odot f_2\left\{g_2\left(S_0\right) \odot f_1\left[g_1\left(S_0\right) \odot f_0\left(S_0\right)\right]\right\}, \tag*{(6)}

where ⊙\odot denotes the Hadamard (element-wise) product.

To illustrate why this structure constitutes a continued function, we draw an analogy with Padé approximants, which can be expressed as continued fractions—a special case of continued functions.

In physics and applied mathematics, perturbative series naturally express physical interactions to arbitrary order. However, these series have a fundamental dilemma: truncating at low order fails to capture the complete physics, while including higher-order terms typically leads to divergence. Renormalization provides the solution to this problem, with Padé approximants being the most widely used numerical scheme for transforming divergent series into convergent expressions. To match a perturbative series f(x)∼c0+c1x+c2x2+⋯f(x) \sim c_0 + c_1x + c_2x^2 + \cdots with a Padé approximant PM(x)QM(x)\frac{P_M(x)}{Q_M(x)}, one equates truncated c0+c1x+c2x2+⋯c_0 + c_1x + c_2x^2 + \cdots with the rational function PM(x)QM(x)\frac{P_M(x)}{Q_M(x)}, then deals with the algebraic equation

(c0+c1x+c2x2+⋯ )QM(x)=PM(x).(c_0 + c_1x + c_2x^2 + \cdots)Q_M(x) = P_M(x).

MM is the order of the polynomial. The approximation accuracy typically improves as M→∞M \to\infty. Examples of continued fraction:

  • Exponential function:

ex∼1+11!x+12!x2+13!x3+14!x4+⋯=1+x1−x2+x−2x3+x−3x4+x−⋱(7)e^x \sim1+\frac{1}{1!}x+\frac{1}{2!}x^2+\frac{1}{3!}x^3+\frac{1}{4!}x^4+\cdots=1+\frac{x}{1-\frac{x}{2+x-\frac{2x}{3+x-\frac{3x}{4+x-\ddots}}}} \tag*{(7)}
  • Natural logarithm:

log⁡(1+x)∼x−12x2+13x3−14x4+⋯=x1+12x2+12x3+22x4+⋱(8)\log(1+x)\sim x-\frac{1}{2}x^2+\frac{1}{3}x^3-\frac{1}{4}x^4+\cdots=\frac{x}{1+\frac{1^2x}{2+\frac{1^2x}{3+\frac{2^2x}{4+\ddots}}}} \tag*{(8)}

The Padé approximant for exe^x in Eq. (7) has everyday application throughout numerical analysis, since this is how matrix exponential is computed in various programming languages [1], for example, the expm in Python, Mathematica, and MATLAB. In Eq. (8), the perturbative series diverges for x≥1x\ge1, however, the Padé approximant always converges and captures well the value of log⁡(1+x)\log(1+x). The continued fraction can be naturally generalized, by replacing the inverse operation with other nonlinear functions. Examples of continued function [2]:

  • CN(z)=c0exp⁡(c1zexp⁡(c2z⋯exp⁡(cNz)))C_N(z)=c_0\exp(c_1z\exp(c_2z\cdots\exp(c_Nz))),

  • CN(z)=c01+c1z1+c2z1+⋯cN−1z1+cNzC_N(z)=c_0\sqrt{1+c_1z\sqrt{1+c_2z\sqrt{1+\cdots c_{N-1}z\sqrt{1+c_Nz}}}},

  • CN(z)=log⁡(c0+zlog⁡(c1+zlog⁡(c2+⋯zlog⁡(cN+z))))C_N(z)=\log(c_0+z\log(c_1+z\log(c_2+\cdots z\log(c_N+z)))).

To see why EFNN is a continued function, we first turn to a simpler case by requiring S0∈RS_0\in\mathbb{R}. We let fj(z)=dj+cjbj+Ajzf_j(z)=d_j+\frac{c_j}{b_j+A_jz}, then S3S_3 in Eq. (3) becomes:

S3=g3(S0)d2+g3(S0)c2b2+A2g2(S0)d1+A2g2(S0)c1b1+A1g1(S0)d0+A1g1(S0)c0b0+A0S0(9)S_3=g_3(S_0)d_2+\frac{g_3(S_0)c_2}{b_2+A_2g_2(S_0)d_1+\frac{A_2g_2(S_0)c_1}{b_1+A_1g_1(S_0)d_0+\frac{A_1g_1(S_0)c_0}{b_0+A_0S_0}}} \tag*{(9)}

Since S0S_0 is actually a vector, and rational functions always introduce fictitious poles, we would like to replace the inverse function with tanh, using Eq. (2). Then the neural network implementation with tanh then gives us:

S3=g3(S0)⊙[d2+c2tanh⁡(b2+A2g2(S0)⊙[d1+c1tanh⁡(b1+A1g1(S0)⊙[d0+c0tanh⁡(b0+A0S0)])])].(10)S_3=g_3(S_0)\odot\left[d_2+c_2\tanh\left(b_2+A_2g_2(S_0)\odot\left[d_1+c_1\tanh\left(b_1+A_1g_1(S_0)\odot\left[d_0+c_0\tanh\left(b_0+A_0S_0\right)\right]\right)\right]\right)\right]. \tag*{(10)}

And we have fully demonstrated that S3S_3 is a continued function.

Note that in Eq. (10), S3S_3 is vector-valued, each component of S3S_3, which is a finite value, contains infinite number of perturbation terms, as shown in Eq. (11).

∑n1,n2,⋯ ,nN=0∞an1n2⋯nN[S0(1)]n1[S0(2)]n2⋯[S0(N)]nN.(11)\sum_{n_1,n_2,\cdots,n_N=0}^{\infty} a_{n_1n_2\cdots n_N}[S_0^{(1)}]^{n_1}[S_0^{(2)}]^{n_2}\cdots[S_0^{(N)}]^{n_N}. \tag*{(11)}

Why ResNet, DenseNet, and DNN are not continued function

For DNN, as shown in Fig. 2(c), the mapping S0↦F1↦S1↦F2↦⋯↦S3=g3(f2(g2(f1(g1f0(S0))))))S_0 \mapsto F_1 \mapsto S_1 \mapsto F_2 \mapsto\cdots\mapsto S_3 = g_3(f_2(g_2(f_1(g_1f_0(S_0)))))), does not contain any skip connections, as required by continued function. Therefore it is apparent that DNN is not a continued function.

For ResNet,

Fj=fj(Sj−1),(12)F_j=f_j(S_{j-1}), \tag*{(12)}
Sj=Sj−1⊕Fj, j=1,2,3,(13)S_j=S_{j-1}\oplus F_j,\ j=1,2,3, \tag*{(13)}

⊕\oplus is element-wise addition in Fig. 2(b). Then

S3=S0+f0(S0)+f1[S0+f0(S0)]+f2{S0+f0(S0)+f1[S0+f0(S0)]}.(14)S_3=S_0+f_0(S_0)+f_1\left[S_0+f_0(S_0)\right]+f_2\left\{S_0+f_0(S_0)+f_1\left[S_0+f_0(S_0)\right]\right\}. \tag*{(14)}

fjf_j is nonlinear. Clearly Eq. (14) is not a correct continued function. The correct form would be:

S3=S0+f2[S0+f1(S0+f0(S0))].(15)S_3=S_0+f_2\left[S_0+f_1(S_0+f_0(S_0))\right]. \tag*{(15)}

There are too many redundant terms in Eq. (14). The expression in Eq. (15) has the structure of Fig. 2(a), with ⊙\odot replaced by ⊕\oplus, and gjg_j set to identity.

For DenseNet, as shown in Fig. 2(d), C\mathcal{C} is vector concatenation: for x∈Rax\in\mathbb{R}^{a}, y∈Rby\in\mathbb{R}^{b}, xCy=[xT,yT]Tx\mathcal{C}y=[x^{\mathsf{T}},y^{\mathsf{T}}]^{\mathsf{T}}. We take nonlinear functions fjf_j, the subsequent layer are:

F1=f1(W0S0+b1),(16)F_1=f_1(W_0S_0+b_1), \tag*{(16)}
S1=[S0F1],(17)S_1=\begin{bmatrix}S_0\\F_1\end{bmatrix}, \tag*{(17)}
F2=f2([W10,W11][S0F1]+b1),(18)F_2=f_2\left([W_{10},W_{11}]\begin{bmatrix}S_0\\F_1\end{bmatrix}+b_1\right), \tag*{(18)}
S2=[S0F1F2],(19)S_2=\begin{bmatrix}S_0\\F_1\\F_2\end{bmatrix}, \tag*{(19)}
F3=f3([W20,W21,W22][S0F1F2]+b2),(20)F_3=f_3\left([W_{20},W_{21},W_{22}]\begin{bmatrix}S_0\\F_1\\F_2\end{bmatrix}+b_2\right), \tag*{(20)}
S3=[S0F1F2F3].(21)S_3=\begin{bmatrix}S_0\\F_1\\F_2\\F_3\end{bmatrix}. \tag*{(21)}

The explicit expressions for F1F_1, F2F_2, F3F_3 read

F1=f0(W0S0+b0),(22)F_1=f_0(W_0S_0+b_0), \tag*{(22)}
F2=f1[W10S0+W11f0(W0S0+b0)+b1],(23)F_2=f_1\left[W_{10}S_0+W_{11}f_0\left(W_0S_0+b_0\right)+b_1\right], \tag*{(23)}
F3=f2{W20S0+W21f0(W0S0+b0)+W22f1[W10S0+W11f0(W0S0+b0)+b1]+b2}.(24)F_3=f_2\left\{W_{20}S_0+W_{21}f_0\left(W_0S_0+b_0\right)+W_{22}f_1\left[W_{10}S_0+W_{11}f_0\left(W_0S_0+b_0\right)+b_1\right]+b_2\right\}. \tag*{(24)}

The expression in Eq. (21) is not a correct continued function. The S3S_3 layer concatenates previous layers, a summation of S3S_3’s components gives energy, such complicated function cannot be interpreted, and S0S_0 is not properly renormalized.

From renormalization group equation to EFNN

In the previous section, we have explained that EFNN is a continued function. In this section, we show that the EFNN structure can be derived from a renormalization group (RG) equation, as all continued functions can be.

We follow the arguments in Refs. [3, 4, 5], to derive the EFNN structure from RG equation. The derivation proceeds in 4 major steps: Step 1, we transform the perturbative series using an algebraic prefactor, and obtain exponential renormalization through RG equation. Step 2, we apply nested renormalization to create the exponential self-similar approximant with hierarchical structure. Step 3, we replace unbounded functions with bounded tanh functions while preserving asymptotic properties. Step 4, we generalize from scalar to vector inputs using Hadamard products, yielding the final EFNN architecture.

Step 1: using algebraic transform to renormalize perturbative series through RG equation

We first use scalar S0∈RS_0 \in\mathbb{R}, target function is E(S0)E(S_0), we know its perturbative series up to kk-th order:

p0(S0)=a0,(25)p_0(S_0) = a_0, \tag*{(25)}
p1(S0)=a0+a1S0,(26)p_1(S_0) = a_0 + a_1S_0, \tag*{(26)}
⋮(27)\vdots \tag*{(27)}
pk−1(S0)=a0+a1S0+⋯+ak−1S0k−1,(28)p_{k-1}(S_0) = a_0 + a_1S_0 + \cdots+ a_{k-1}S_0^{k-1}, \tag*{(28)}
pk(S0)=a0+a1S0+⋯+ak−1S0k−1+akS0k.(29)p_k(S_0) = a_0 + a_1S_0 + \cdots+ a_{k-1}S_0^{k-1} + a_kS_0^k. \tag*{(29)}

We apply an algebraic transform to the above polynomials:

u0=S0θp0(S0)=a0S0θ,(30)u_0 = S_0^\theta p_0(S_0) = a_0S_0^\theta, \tag*{(30)}
u1=S0θp1(S0)=a0S0θ+a1S01+θ,(31)u_1 = S_0^\theta p_1(S_0) = a_0S_0^\theta+ a_1S_0^{1+\theta}, \tag*{(31)}
⋮(32)\vdots \tag*{(32)}
uk−1(S0)=S0θpk−1(S0)=a0S0θ+a1S01+θ+⋯+ak−1S0θ+k−1(33)u_{k-1}(S_0) = S_0^\theta p_{k-1}(S_0) = a_0S_0^\theta+ a_1S_0^{1+\theta} + \cdots+ a_{k-1}S_0^{\theta+k-1} \tag*{(33)}
uk(S0)=S0θpk(S0)=a0S0θ+a1S01+θ+⋯+ak−1S0θ+k−1+akS0θ+k.(34)u_k(S_0) = S_0^\theta p_k(S_0) = a_0S_0^\theta+ a_1S_0^{1+\theta} + \cdots+ a_{k-1}S_0^{\theta+k-1} + a_kS_0^{\theta+k}. \tag*{(34)}

We embed the discrete functions {uj(S0):j∈N}\{u_j(S_0) : j \in\mathbb{N}\} into {u(t,S0):t≥0}\{u(t,S_0) : t \ge0\}, with uj(S0)=u(j,S0)u_j(S_0) = u(j,S_0), since continuous variable is easier to deal with. tt is interpreted as time. We let u(0,S0)=fu(0,S_0) = f, and further denote u(t,S0)u(t,S_0) as u(t,S0,f)u(t,S_0,f), to show that the initial value of the flow is ff, while the variable in physical space is S0S_0. We assume that uu satisfies the self-similar equation:

u(t2+t1,S0,f)=u(t2,S0,u(t1,S0,f)),(35)u(t_2+t_1,S_0,f) = u(t_2,S_0,u(t_1,S_0,f)), \tag*{(35)}

because we hope to find the fixed point solution u∗u^*:

u(t,S0,u∗(S0))=u∗(S0),(36)u(t,S_0,u^*(S_0)) = u^*(S_0), \tag*{(36)}

to guarantee the best convergence when using variable S0S_0 [6], i.e., in the physical space. u∗(S0)u^*(S_0) does not depend on tt, because it is already a fixed point solution. u∗(S0)u^*(S_0) does not depend on initial value ff, because we hope that u∗u^* is an attractor. Eq. (32) leads to the Lie equation:

∂tu(t,S0,f)=v(u(t,S0,f)),(37)\partial_tu(t,S_0,f) = v(u(t,S_0,f)), \tag*{(37)}

the velocity is given by

v(u(t,S0,f))=∂ϵu(ϵ,S0,u(t,S0,f))∣ϵ=0.(38)v(u(t,S_0,f)) = \left.\partial_\epsilon u(\epsilon,S_0,u(t,S_0,f))\right|_{\epsilon=0}. \tag*{(38)}

Lie equation is exactly RG equation, when tt is scaled as t=log⁡τt = \log\tau:

∂u∂log⁡τ=v(u).(39)\frac{\partial u}{\partial\log\tau} = v(u). \tag*{(39)}

We can’t obtain the fixed point solution u∗u^* exactly, and need to compute the so called kk-th order quasi-fixed point solution. We use the Euler scheme to approximate the velocity at t=kt=k:

v(u(k,S0,f))≃u(k,S0,f)−u(k−1,S0,f)=akS0θ+k.(40)v(u(k,S_0,f)) \simeq u(k,S_0,f)-u(k-1,S_0,f)=a_k S_0^{\theta+k}. \tag*{(40)}

We assume that starting from uk−1u_{k-1}, it takes time TkT_k to reach the kk-th order quasi-fixed point uk∗u_k^*, integrating the Lie equation Eq. (37), we obtain:

∫uk−1uk∗duakξθ+k=Tk.(41)\int_{u_{k-1}}^{u_k^*}\frac{\mathrm{d}u}{a_k \xi^{\theta+k}}=T_k. \tag*{(41)}

The dummy variable ξ\xi is in physical space, starting with S0S_0 for lower limit uk−1(S0,f)u_{k-1}(S_0,f). We choose the renormalization map from physical space to function space as algebraic:

u=bξθ,(42)u=b\xi^\theta, \tag*{(42)}

then

uk∗=(uk−1−kθ−kθakb1+kθTk)−θk.(43)u_k^*=\left(u_{k-1}^{-\frac{k}{\theta}}-\frac{k}{\theta}\frac{a_k}{b^{1+\frac{k}{\theta}}}T_k\right)^{-\frac{\theta}{k}}. \tag*{(43)}

We multiply the above equation by S0−θS_0^{-\theta} to scale back, and obtain a renormalized value for E(S0)E(S_0):

Ek∗(S0)=S0−θuk∗=pk−1(S0)[1−kθakb1+kθTkS0kpk−1kθ(S0)]−θk.(44)\begin{aligned} E_k^*(S_0)=S_0^{-\theta}u_k^* \\ &= p_{k-1}(S_0)\left[1-\frac{k}{\theta}\frac{a_k}{b^{1+\frac{k}{\theta}}}T_k S_0^k p_{k-1}^{\frac{k}{\theta}}(S_0)\right]^{-\frac{\theta}{k}}. \tag*{(44)} \end{aligned}

Starting from Eq. (44), by tuning values of θ\theta, bb, TkT_k, we can arrive at several continued function approximants, including Padé. Here, we let θ→∞\theta\to\infty, and obtain

Ek∗(S0)=pk−1(S0)exp⁡(akbTkS0k).(45)E_k^*(S_0)=p_{k-1}(S_0)\exp\left(\frac{a_k}{b}T_kS_0^k\right). \tag*{(45)}

This is to say, this one-step renormalization is a map:

pk(S0)=a0+a1S0+⋯+akS0k↦Ek∗(S0)=(a0+a1S0+⋯+ak−1S0k−1)exp⁡(akbTkS0k).(46)p_k(S_0)=a_0+a_1S_0+\cdots+a_kS_0^k \mapsto E_k^*(S_0)=(a_0+a_1S_0+\cdots+a_{k-1}S_0^{k-1})\exp\left(\frac{a_k}{b}T_kS_0^k\right). \tag*{(46)}

We repeat the renormalization process for the polynomial prefactor until only the affine term is left:

pk(S0)↦(a0+a1S0)exp⁡(a2bT2S02+a3bT3S03+⋯+akbTkS0k).(47)p_k(S_0)\mapsto(a_0+a_1S_0)\exp\left(\frac{a_2}{b}T_2S_0^2+\frac{a_3}{b}T_3S_0^3+\cdots+\frac{a_k}{b}T_kS_0^k\right). \tag*{(47)}

Step 2: using nested renormalization to obtain exponential self-similar approximant

We use the exponential renormalization to renormalize the polynomial in the exponential part in Eq. (47), and always leave a linear term as a prefactor, we obtain the exponential self-similar approximant:

pk(S0)↦(a0+a1S0)exp⁡(b1S0exp⁡(b2S0⋯exp⁡(bnkS0))).(48)p_k(S_0)\mapsto(a_0+a_1S_0)\exp(b_1S_0\exp(b_2S_0\cdots\exp(b_{n_k}S_0))). \tag*{(48)}

nkn_k is an integer depending on kk. One often performs a perturbation analysis for the exponential self-similar approximant with the perturbative series to obtain the unknown coefficients b1b_1, b2b_2, …\ldots, bnkb_{n_k}. Note that one may also include the linear term a1S0a_1S_0 in the exponential, we have the freedom of not taking this approach, and we have numerically tested that this approach does not work well.

Step 3: further renormalization for exponential self-similar approximant

Since linear functions and exponential functions are not bounded, to achieve better convergence, we need to further renormalize the nested exponential in Eq. (48). Since we do not use order by order matching to determine the unknown coefficients, we can set nkn_k with a given integer. We take nk=3n_k=3 as an example. To further renormalize, we proceed in 5 substeps. We denote the exponential continued function as

Φ1(S0)=(α3S0+κ3)exp⁡(α2S0exp⁡(α1S0exp⁡(A0S0))).(49)\Phi_1(S_0)=(\alpha_3S_0+\kappa_3)\exp\left(\alpha_2S_0\exp\left(\alpha_1S_0\exp(A_0S_0)\right)\right). \tag*{(49)}

First, we scale and shift the exp function in Φ1(S0)\Phi_1(S_0):

Φ1(S0)↦Φ2(S0)=(α3S0+κ3)[d2′+c2exp⁡(b2+α2S0[d1′+c1exp⁡(b1+α1S0[d0′+c0exp⁡(A0S0)])])].(50)\begin{aligned} \Phi_1(S_0)&\mapsto\Phi_2(S_0)\\ &=(\alpha_3S_0+\kappa_3)\left[d_2'+c_2\exp\left(b_2+\alpha_2S_0\left[d_1'+c_1\exp\left(b_1+\alpha_1S_0\left[d_0'+c_0\exp(A_0S_0)\right]\right)\right]\right)\right]. \tag*{(50)} \end{aligned}

Second, we shift the origin of S0S_0 in Φ2(S0)\Phi_2(S_0):

Φ2(S0)↦Φ3(S0)=(α3S0+θ3)[d2′+c2exp⁡(b2+(α2S0+θ2)[d1′+c1exp⁡(b1+(α1S0+θ1)[d0′+c0exp⁡(b0+A0S0)])])].(51)\begin{aligned} \Phi_2(S_0)&\mapsto\Phi_3(S_0)\\ &=(\alpha_3S_0+\theta_3)\left[d_2'+c_2\exp\left(b_2+(\alpha_2S_0+\theta_2)\left[d_1'+c_1\exp\left(b_1+(\alpha_1S_0+\theta_1)\left[d_0'+c_0\exp\left(b_0+A_0S_0\right)\right]\right)\right]\right)\right]. \tag*{(51)} \end{aligned}

Third, renormalize exp by tanh. We know that 1+tanh⁡(y)∼exp⁡(y)1+\tanh(y)\sim\exp(y), when ∣y∣\lvert y\rvert is small, and tanh⁡(y)\tanh(y) is bounded on R\mathbb{R}. Therefore we replace exp⁡(y)\exp(y) by 1+tanh⁡(y)1+\tanh(y):

Φ3(S0)↦Φ4(S0)=(α3S0+θ3)[d2+c2tanh⁡(b2+(α2S0+θ2)[d1+c1tanh⁡(b1+(α1S0+θ1)[d0+c0tanh⁡(b0+A0S0)])])].(52)\begin{aligned} \Phi_3(S_0)&\mapsto\Phi_4(S_0)\\ &=(\alpha_3S_0+\theta_3)\left[d_2+c_2\tanh\left(b_2+(\alpha_2S_0+\theta_2)\left[d_1+c_1\tanh\left(b_1+(\alpha_1S_0+\theta_1)\left[d_0+c_0\tanh\left(b_0+A_0S_0\right)\right]\right)\right]\right)\right]. \tag*{(52)} \end{aligned}

Fourth, renormalize the linear prefactors before tanh. Since y∼tanh⁡(y)y\sim\tanh(y), when ∣y∣\lvert y\rvert is small, we perform the following renormalizations:

αjS0+θj↦gj(S0)=βjtanh⁡(γjS0+δj)+φj,j=1,2,3.(53)\alpha_jS_0+\theta_j\mapsto g_j(S_0)=\beta_j\tanh(\gamma_jS_0+\delta_j)+\varphi_j,\quad j=1,2,3. \tag*{(53)}

Then

Φ4(S0)↦Φ5(S0)=g3(S0)[d2+c2tanh⁡(b2+g2(S0)[d1+c1tanh⁡(b1+g1(S0)[d0+c0tanh⁡(b0+A0S0)])])].(54)\begin{aligned} \Phi_4(S_0)&\mapsto\Phi_5(S_0)\\ &=g_3(S_0)\left[d_2+c_2\tanh\left(b_2+g_2(S_0)\left[d_1+c_1\tanh\left(b_1+g_1(S_0)\left[d_0+c_0\tanh\left(b_0+A_0S_0\right)\right]\right)\right]\right)\right]. \tag*{(54)} \end{aligned}

Fifth, rescale g1(S0)g_1(S_0), g2(S0)g_2(S_0):

Φ5(S0)↦Φ6(S0)=g3(S0)[d2+c2tanh⁡(b2+A2g2(S0)[d1+c1tanh⁡(b1+A1g1(S0)[d0+c0tanh⁡(b0+A0S0)])])].(55)\begin{aligned} \Phi_5(S_0)&\mapsto\Phi_6(S_0)\\ &=g_3(S_0)\left[d_2+c_2\tanh\left(b_2+A_2g_2(S_0)\left[d_1+c_1\tanh\left(b_1+A_1g_1(S_0)\left[d_0+c_0\tanh\left(b_0+A_0S_0\right)\right]\right)\right]\right)\right]. \tag*{(55)} \end{aligned}

Step 4: generalization to vector case

The above derivations take S0∈RS_0\in\mathbb{R}. For a spin system, S0S_0 is vector-valued. We generalize Φ6(S0)\Phi_6(S_0), let S0∈RNS_0\in\mathbb{R}^N, and transform scalar product to Hadamard product ⊙\odot, and let bk∈RNb_k\in\mathbb{R}^N, dk∈RNd_k\in\mathbb{R}^N, Ak∈RN×NA_k\in\mathbb{R}^{N\times N}, ck∈RN×Nc_k\in\mathbb{R}^{N\times N}, k=0,1,2k=0,1,2. For gj(S0)g_j(S_0), j=1,2,3j=1,2,3, we let γj∈RN×N\gamma_j \in\mathbb{R}^{N \times N}, βj∈RN×N\beta_j \in\mathbb{R}^{N \times N}, δj∈RN\delta_j \in\mathbb{R}^{N}, φj∈RN\varphi_j \in\mathbb{R}^{N}. Then

Φ6(S0)=g3(S0)⊙[d2+c2tanh⁡(b2+A2g2(S0)⊙[d1+c1tanh⁡(b1+A1g1(S0)⊙[d0+c0tanh⁡(b0+A0S0)])])].(56)\Phi_6(S_0)=g_3(S_0)\odot\left[d_2+c_2\tanh\left(b_2+A_2g_2(S_0)\odot\left[d_1+c_1\tanh\left(b_1+A_1g_1(S_0)\odot\left[d_0+c_0\tanh\left(b_0+A_0S_0\right)\right]\right)\right]\right)\right]. \tag*{(56)}

The above Φ6(S0)\Phi_6(S_0) is just the S3S_3 in Eq. (10). We need one more step to obtain the energy, i.e., summing up all components of Φ6(S0)\Phi_6(S_0) to get a real scalar as the predicted energy.

Symmetrization

In the main text, the symmetrization layer is given by

Tj(S0)=∑k=x,y,zconv⁡(S0,k)⊙conv⁡(S0,k),j=0,1,…,n.(57)T_j(S_0)=\sum_{k=x,y,z}\operatorname{conv}(S_{0,k})\odot\operatorname{conv}(S_{0,k}),\quad j=0,1,\ldots,n. \tag*{(57)}

We choose to use 2 different matrices as the filters for the 2 convolutional operations in Tj(S0)T_j(S_0), and rewrite Tj(S0)T_j(S_0) as

Tj(S0)=∑k=x,y,zconv⁡2j(S0,k)⊙conv⁡2j+1(S0,k).(58)T_j(S_0)=\sum_{k=x,y,z}\operatorname{conv}_{2j}(S_{0,k})\odot\operatorname{conv}_{2j+1}(S_{0,k}). \tag*{(58)}

For conv⁡q\operatorname{conv}_q, the convolution filter is Wq∈RC×R×SW_q\in\mathbb{R}^{C\times R\times S}, CC is the output channel number. The [a,b,c][a,b,c]-th element of the convolution result is

conv⁡q(S0,k)[a,b,c]=∑α=0R−1∑β=0S−1S0,k[b+α,c+β]Wq[a,α,β],(59)\operatorname{conv}_q(S_{0,k})[a,b,c]=\sum_{\alpha=0}^{R-1}\sum_{\beta=0}^{S-1}S_{0,k}[b+\alpha,c+\beta]W_q[a,\alpha,\beta], \tag*{(59)}

where we assume a proper padding scheme is chosen in the above equation. Then

conv⁡2j[a,b,c]=∑α=0R−1∑β=0S−1S0,k[b+α,c+β]W2j[a,α,β],(60)\operatorname{conv}_{2j}[a,b,c]=\sum_{\alpha=0}^{R-1}\sum_{\beta=0}^{S-1}S_{0,k}[b+\alpha,c+\beta]W_{2j}[a,\alpha,\beta], \tag*{(60)}
conv⁡2j+1[a,b,c]=∑γ=0R−1∑δ=0S−1S0,k[b+γ,c+δ]W2j+1[a,γ,δ],(61)\operatorname{conv}_{2j+1}[a,b,c]=\sum_{\gamma=0}^{R-1}\sum_{\delta=0}^{S-1}S_{0,k}[b+\gamma,c+\delta]W_{2j+1}[a,\gamma,\delta], \tag*{(61)}

the [a,b,c][a,b,c]-th element of the symmetrization layer reads

Tj(S0)[a,b,c]=∑k=x,y,zconv⁡2j(S0,k)[a,b,c]⋅conv⁡2j+1(S0,k)[a,b,c]=∑α,β,γ,δW2j[a,α,β]W2j+1[a,γ,δ]∑k=x,y,zS0,k[b+α,c+β]S0,k[b+γ,c+δ].(62)\begin{aligned} T_j(S_0)[a,b,c] &=\sum_{k=x,y,z}\operatorname{conv}_{2j}(S_{0,k})[a,b,c]\cdot\operatorname{conv}_{2j+1}(S_{0,k})[a,b,c]\\ &=\sum_{\alpha,\beta,\gamma,\delta}W_{2j}[a,\alpha,\beta]W_{2j+1}[a,\gamma,\delta]\sum_{k=x,y,z}S_{0,k}[b+\alpha,c+\beta]S_{0,k}[b+\gamma,c+\delta]. \tag*{(62)} \end{aligned}

We let

s1=[S0,x[b+α,c+β]S0,y[b+α,c+β]S0,z[b+α,c+β]].(63)s_1=\begin{bmatrix} S_{0,x}[b+\alpha,c+\beta]\\ S_{0,y}[b+\alpha,c+\beta]\\ S_{0,z}[b+\alpha,c+\beta] \end{bmatrix}. \tag*{(63)}
s2=[S0,x[b+γ,c+δ]S0,y[b+γ,c+δ]S0,z[b+γ,c+δ]].(64)s_2=\begin{bmatrix} S_{0,x}[b+\gamma,c+\delta]\\ S_{0,y}[b+\gamma,c+\delta]\\ S_{0,z}[b+\gamma,c+\delta] \end{bmatrix}. \tag*{(64)}
Image not extracted

Figure 3 (PDF p. 17). (a) ResNet, (b) DenseNet, and (c) DNN adapted for Monte Carlo energy fitting of quantum double exchange model. ⊕\oplus is element-wise addition, while CC is concatenation.

Then

s1Ts2=∑k=x,y,zS0,k[b+α,c+β]S0,k[b+γ,c+δ].(65)s_1^{\mathsf{T}}s_2=\sum_{k=x,y,z} S_{0,k}[b+\alpha,c+\beta]S_{0,k}[b+\gamma,c+\delta]. \tag*{(65)}

For a matrix U∈O(3)U \in O(3),

(Us1)TUs2=s1TUTUs2=s1Ts2.(66)(Us_1)^{\mathsf{T}}Us_2=s_1^{\mathsf{T}}U^{\mathsf{T}}Us_2 =s_1^{\mathsf{T}}s_2. \tag*{(66)}

As a result, under any operation U∈O(3)U \in O(3), the value Tj[a,b,c]T_j[a,b,c] remains invariant.

ResNet, DenseNet, DNN adapted for Monte Carlo energy fitting of quantum double exchange model

In this section, we present the structures of ResNet, DenseNet, and DNN, adapted for Monte Carlo energy fitting of quantum double exchange model.

As shown in Fig. 3(a) and Fig. 3(b), for ResNet and DenseNet, we add 2 symmetrization layers T0T_0 and T1T_1 to obtain O(3)O(3) invariant results. For DNN in Fig. 3(c), we use 1 symmetrization layer T0T_0, as no skip connection is used for DNN. How the nonlinear layers are computed, how the symmetrization is performed, and how the final energy is computed are similar to Fig. 3 in the main text.

DQMC and the speed-up by EFNN

In this section, we explain the computational bottleneck problem in Determinant Quantum Monte Carlo (DQMC), and report the computational speed-up achieved by EFNN.

Outline of DQMC computation in 2D quantum double exchange model

For this quantum double exchange model, the DQMC computation has 3 major steps [7, 8].

Converting the Hamiltonian to quadratic form

The 2D quantum double exchange model is a lattice system describing electronic hopping interactions and on-site coupling between itinerant electrons and localized classical spins. In this framework, the classical spins act as a random field. The Hamiltonian reads (with chemical potential part)

H(S)=−t∑⟨i,j⟩,α(ciα†cjα+cjα†ciα)−J2∑i,α,βs⃗i⋅ciα†σ⃗αβciβ−μ∑iαniα,(67)H(S)=-t\sum_{\langle i,j\rangle,\alpha}\left(c_{i\alpha}^{\dagger}c_{j\alpha}+c_{j\alpha}^{\dagger}c_{i\alpha}\right)-\frac{J}{2}\sum_{i,\alpha,\beta}\vec{s}_{i}\cdot c_{i\alpha}^{\dagger}\vec{\sigma}_{\alpha\beta}c_{i\beta}-\mu\sum_{i\alpha}n_{i\alpha}, \tag*{(67)}

−t∈R-t \in\mathbb{R} is the hopping amplitude, −J2-\frac{J}{2} is the on-site interaction strength, s⃗i∈S2\vec{s}_{i} \in S^{2} is a classical spin vector. σx\sigma^{x}, σy\sigma^{y}, σz\sigma^{z} are Pauli matrices. α\alpha, β\beta denote spin components ↑\uparrow and ↓\downarrow.

We first need to expand the site index ii to double index, the double index read [a,b][a,b], a,b=0,1,…,N−1a,b=0,1,\ldots,N-1, denoting a lattice point on a square. Now a creation operator reads cab↑†c_{ab\uparrow}^{\dagger} or cab↓†c_{ab\downarrow}^{\dagger}.

We group the creation operators into a row vector. The row vector has two parts,

c†=[c↑†,c↓†],(68)c^{\dagger}=[c_{\uparrow}^{\dagger},c_{\downarrow}^{\dagger}], \tag*{(68)}

the first part c↑†c_{\uparrow}^{\dagger} is a row of creation operators with spin ↑\uparrow, the explicit expression reads

c↑†=[c00↑†,c01↑†,…,c0,N−1,↑†,c10↑†,c11↑†,…,c1,N−1,↑†,…,cN−1,0,↑†,cN−1,1,↑†,…,cN−1,N−1,↑†].(69)c_{\uparrow}^{\dagger}=[c_{00\uparrow}^{\dagger},c_{01\uparrow}^{\dagger},\ldots,c_{0,N-1,\uparrow}^{\dagger},c_{10\uparrow}^{\dagger},c_{11\uparrow}^{\dagger},\ldots,c_{1,N-1,\uparrow}^{\dagger},\ldots,c_{N-1,0,\uparrow}^{\dagger},c_{N-1,1,\uparrow}^{\dagger},\ldots,c_{N-1,N-1,\uparrow}^{\dagger}]. \tag*{(69)}

The creation operators in Eq. (69) are arranged in row-major order. Similarly, we arrange ↓\downarrow creation operators in row-major order:

c↓†=[c00↓†,c01↓†,…,c0,N−1,↓†,c10↓†,c11↓†,…,c1,N−1,↓†,…,cN−1,0,↓†,cN−1,1,↓†,…,cN−1,N−1,↓†].(70)c_{\downarrow}^{\dagger}=[c_{00\downarrow}^{\dagger},c_{01\downarrow}^{\dagger},\ldots,c_{0,N-1,\downarrow}^{\dagger},c_{10\downarrow}^{\dagger},c_{11\downarrow}^{\dagger},\ldots,c_{1,N-1,\downarrow}^{\dagger},\ldots,c_{N-1,0,\downarrow}^{\dagger},c_{N-1,1,\downarrow}^{\dagger},\ldots,c_{N-1,N-1,\downarrow}^{\dagger}]. \tag*{(70)}

The Hamiltonian in Eq. (67) is expressed in a quadratic form:

H(S)=c†Γ(S)c,(71)H(S)=c^{\dagger}\Gamma(S)c, \tag*{(71)}

Hermitian matrix Γ(S)∈C2N2×2N2\Gamma(S)\in\mathbb{C}^{2N^{2}\times2N^{2}},

Γ(S)=[Tx+Ty00Tx+Ty]−J2[SzSx−iSySx+iSy−Sz]+[−μI00−μI].(72)\Gamma(S)= \begin{bmatrix} T_{x}+T_{y} & 0\\ 0 & T_{x}+T_{y} \end{bmatrix} -\frac{J}{2} \begin{bmatrix} S^{z} & S^{x}-iS^{y}\\ S^{x}+iS^{y} & -S^{z} \end{bmatrix} + \begin{bmatrix} -\mu I & 0\\ 0 & -\mu I \end{bmatrix}. \tag*{(72)}

TxT_{x} and TyT_{y} are hopping matrices in xx and yy directions. For Tx∈RN2×N2T_{x}\in\mathbb{R}^{N^{2}\times N^{2}}, the [rN+s,(r+1)N+s][rN+s,(r+1)N+s]-th element is −t-t, the [(r+1)N+s,rN+s][(r+1)N+s,rN+s]-th element is −t-t, the rest are 0. The boundary condition is PBC, the rr, ss indices are understood as r mod⁡Nr\ \operatorname{mod} N, s mod⁡Ns\ \operatorname{mod} N. For hopping matrix Ty∈RN2×N2T_{y}\in\mathbb{R}^{N^{2}\times N^{2}}, the [rN+s,rN+(s+1)][rN+s,rN+(s+1)]-th element and the [rN+(s+1),rN+s][rN+(s+1),rN+s]-th element are −t-t, the rest are 0. The boundary condition is PBC, the rr, ss indices are understood as r mod⁡Nr\ \operatorname{mod} N, s mod⁡Ns\ \operatorname{mod} N.

Matrix SxS^{x} contains all the sabxs_{ab}^{x} components along the diagonal:

Sx=diag⁡(s00x,s01x,…,s0,N−1x,s10x,s11x,…,s1,N−1x,…,sN−1,0x,sN−1,1x,…,sN−1,N−1x).(73)S^{x}=\operatorname{diag}(s_{00}^{x},s_{01}^{x},\ldots,s_{0,N-1}^{x},s_{10}^{x},s_{11}^{x},\ldots,s_{1,N-1}^{x},\ldots,s_{N-1,0}^{x},s_{N-1,1}^{x},\ldots,s_{N-1,N-1}^{x}). \tag*{(73)}

Matrix SyS^{y} contains all the sabys_{ab}^{y} components along the diagonal:

Sy=diag⁡(s00y,s01y,…,s0,N−1y,s10y,s11y,…,s1,N−1y,…,sN−1,0y,sN−1,1y,…,sN−1,N−1y).(74)S^{y}=\operatorname{diag}(s_{00}^{y},s_{01}^{y},\ldots,s_{0,N-1}^{y},s_{10}^{y},s_{11}^{y},\ldots,s_{1,N-1}^{y},\ldots,s_{N-1,0}^{y},s_{N-1,1}^{y},\ldots,s_{N-1,N-1}^{y}). \tag*{(74)}

Matrix SzS^z contains all the sabzs^z_{ab} components along the diagonal:

Sz=diag⁡(s00z,s01z,…,s0,N−1z,s10z,s11z,…,s1,N−1z,…,sN−1,0z,sN−1,1z,…,sN−1,N−1z).(75)S^z = \operatorname{diag}(s^z_{00},s^z_{01},\ldots,s^z_{0,N-1},s^z_{10},s^z_{11},\ldots,s^z_{1,N-1},\ldots,s^z_{N-1,0},s^z_{N-1,1},\ldots,s^z_{N-1,N-1}). \tag*{(75)}

Convert tr to det

In this step, with the quadratic form in Eq. (71), the trace is converted to det, following the DQMC scheme:

Z(S)=∑Str⁡e−βH(S)=∑Str⁡e−βc†Γ(S)c=∑Sdet⁡[I+e−βΓ(S)].(76)\begin{aligned} Z(S) &= \sum_S \operatorname{tr}e^{-\beta H(S)} \\ &= \sum_S \operatorname{tr}e^{-\beta c^\dagger\Gamma(S)c} \\ &= \sum_S \det[I+e^{-\beta\Gamma(S)}]. \tag*{(76)} \end{aligned}

In this step, the trace of a 4N2×4N24N^2 \times4N^2 matrix is converted to the determinant of a 2N2×2N22N^2 \times2N^2 matrix, this is the quantum-to-classical mapping in DQMC.

More precisely, the partition function is

Z(S)=∫{s⃗ab∈S2:a,b=0,1,…,N−1}tr⁡e−βH(S)dS,(77)Z(S)=\int_{\{\vec{s}_{ab}\in S^2:a,b=0,1,\ldots,N-1\}}\operatorname{tr}e^{-\beta H(S)}dS, \tag*{(77)}

since all the classical continuous spins have norm 1. Eq. (76) can replace Eq. (77), only differing by a constant factor when the data points are dense and even enough.

Classical Monte Carlo using determinants

We use the determinants in Eq. (76) as statistical weight for Monte Carlo sampling. Determinants in Eq. (76) have the nice property that they are all strictly positive, because they can be written as

det⁡[I+e−βΓ(S)]=∏j=02N2−1[1+e−βEj(S)]>0,(78)\det[I+e^{-\beta\Gamma(S)}]=\prod_{j=0}^{2N^2-1}[1+e^{-\beta E_j(S)}]>0, \tag*{(78)}

{Ej(S)}\{E_j(S)\} are the eigenvalues of Hermitian matrix Γ(S)\Gamma(S). In each Monte Carlo step, we perform update on one component of one classical spin vector, among N2N^2 such 3D vectors in S2S^2. Evaluating Eq. (78) requires Exact Diagonalization (ED), for such 2N2×2N22N^2 \times2N^2 matrix, one evaluation requires complexity O((2N2)3)=O(8N6)O((2N^2)^3)=O(8N^6). Note that evaluating the det in left hand side of Eq. (78) has exactly the same complexity. There are 3N23N^2 scalar components in SS in total, a sweep to update every component has complexity 3N2⋅O(8N6)∼O(24N8)3N^2\cdot O(8N^6)\sim O(24N^8). In order to obtain convergent results, we typically need 500 effective samples, then the total computational cost is

500×(autocorrelation length)×(sweep number)×O(24N8).(79)500\times(\text{autocorrelation length})\times(\text{sweep number})\times O(24N^8). \tag*{(79)}

As NN increases, with computational complexity in Eq. (79), the computation becomes prohibitive for classical Monte Carlo very quickly. We use EFNN to address the computational bottleneck problem by training EFNN using exact value of Eq. (78), given SS. We make one more transformation, and define a Monte Carlo energy EMC(S)E_{\mathrm{MC}}(S),

e−βEMC(S)=det⁡[I+e−βΓ(S)]=∏j=02N2−1[1+e−βEj(S)],(80)\begin{aligned} e^{-\beta E_{\mathrm{MC}}(S)} &= \det[I+e^{-\beta\Gamma(S)}] \\ &= \prod_{j=0}^{2N^2-1}[1+e^{-\beta E_j(S)}], \tag*{(80)} \end{aligned}

then

EMC(S)=−T∑j=02N2−1log⁡[1+e−1TEj(S)].(81)E_{\mathrm{MC}}(S)=-T\sum_{j=0}^{2N^2-1}\log[1+e^{-\frac{1}{T}E_j(S)}]. \tag*{(81)}

EMC(S)E_{\mathrm{MC}}(S) is formally a free energy. We use ED (Exact Diagonalization) to generate [S,EMC(S)][S, E_{\mathrm{MC}}(S)] data as training set to EFNN. ED has the same complexity as determinant computation.

NN10152025303540
EFNN time/s4.56×10−34.56 \times10^{-3}1.6×10−21.6 \times10^{-2}1.62×10−21.62 \times10^{-2}1.40×10−21.40 \times10^{-2}7.08×10−37.08 \times10^{-3}9.38×10−39.38 \times10^{-3}7.59×10−37.59 \times10^{-3}
ED time/s6.24×10−36.24 \times10^{-3}3.11×10−23.11 \times10^{-2}1.17×10−11.17 \times10^{-1}4.82×10−14.82 \times10^{-1}1.461.463.563.567.767.76

Table 1. Average computational time for EFNN inference, C=3C = 3, 3 layers, and ED, quantum double exchange model

Image not extracted

Figure 4 (PDF p. 20). Average computational time for EFNN inference, C=3C = 3, 3 layers, and ED. The y-axis has log scale.

We report the average computation times of EFNN and ED in Table I, and make a plot in Fig. 4, for comparison. While the EFNN model was trained on a supercomputer, this benchmark compares the inference time for a single EMC(S)E_{\mathrm{MC}}(S) value given a configuration SS against the corresponding ED calculation, both performed on an Alienware m15 R12 laptop. As shown in Table I and the green line in Fig. 4, the EFNN inference time remains approximately constant at order 10−2 s10^{-2}\,\mathrm{s} when the lattice size N×NN \times N grows from 10×1010 \times10 to 40×4040 \times40. Conversely, as shown by the red line, the ED computational time scales polynomially with NN, rapidly reaching approximately 10 s10\,\mathrm{s} for N=40N = 40. This scaling behavior makes ED computationally prohibitive for Monte Carlo sampling. Notably, for N=40N = 40, the EFNN achieves a speedup of approximately 10310^3.

Details of inference error and data distribution, quantum double exchange model

In this section, we report the error of EFNN prediction, plot the distributions of the test data (ground truth) and the EFNN predictions, for the quantum double exchange model.

We report the absolute value of mean of EMCE_{\mathrm{MC}} for test set, RMSE of prediction by EFNN, and relative error in Table II, with C=3C = 3, layer number =3= 3, in quantum double exchange model. The relative error is defined as

NN10152025303540
∣mean of EMC∣|\mathrm{mean\ of\ } E_{\mathrm{MC}}|51.6508116.1850206.5729322.7573464.7841632.5962826.2887
RMSE0.05840.09890.13400.17140.21180.25490.3003
relative error1.13×10−31.13 \times10^{-3}8.51×10−48.51 \times10^{-4}6.49×10−46.49 \times10^{-4}5.31×10−45.31 \times10^{-4}4.56×10−44.56 \times10^{-4}4.03×10−44.03 \times10^{-4}3.63×10−43.63 \times10^{-4}

Table II. Performance metrics for EFNN, C=3C = 3, 3 layers, quantum double exchange model

RMSE∣mean of EMC∣(82)\frac{\mathrm{RMSE}}{\lvert\mathrm{mean\ of\ } E_{\mathrm{MC}}\rvert} \tag*{(82)}

As shown in Table II, even though the model is trained on a 10×1010 \times10 lattice under boundary condition PBC, when applied to larger lattices, the RMSE increases only marginally. Moreover, the ∣mean of EMC∣|\mathrm{mean\ of\ } E_{\mathrm{MC}}| increases very quickly with respect to NN. As a result, the difference of the prediction value by EFNN and the ground truth value of EMC(S)E_{\mathrm{MC}}(S), becomes smaller and smaller perturbations as NN increases, demonstrating the accuracy of EFNN.

We present the histograms of test set data (YtestY_{\mathrm{test}}, ground truth value of EMC(S)E_{\mathrm{MC}}(S)) and prediction results computed by EFNN (YpredY_{\mathrm{pred}}, predicted value of EMC(S)E_{\mathrm{MC}}(S)) for the quantum double exchange model in Fig. 5, for N=10,15,20,25,30,35,40N = 10, 15, 20, 25, 30, 35, 40. The blue histograms are for ground truth data YtestY_{\mathrm{test}}, and green histograms are for predictions YpredY_{\mathrm{pred}}. Note that as lattice size increases, the model trained on 10×1010 \times10 lattice still correctly predicts EMC(S)E_{\mathrm{MC}}(S) for larger lattices, up to N=40N = 40, as shown by the overall histogram shapes, and the mean values of YtestY_{\mathrm{test}} and YpredY_{\mathrm{pred}}. Note that as NN increases, the absolute value of mean YtestY_{\mathrm{test}} increases.

Image not extracted

Figure 5 (PDF p. 22). EFNN’s prediction results of quantum double exchange model, with C=3C = 3 and 3 layers: histogram of YtestY_{\mathrm{test}} and YpredY_{\mathrm{pred}}, for N=10N = 10, 15, 20, 25, 30, 35, 40.

References

  1. [1]A. H. Al-Mohy and N. J. Higham, A New Scaling and Squaring Algorithm for the Matrix Exponential, SIAM Journal on Matrix Analysis and Applications 31, 970 (2010).DOI
  2. [2]C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists and Engineers I, 1st ed. (Springer New York, New York, NY, 1999).DOI
  3. [3]V. Yukalov and E. Yukalova, Higher Orders of Self-Similar Approximations for Thermodynamic Potentials, Physica A: Statistical Mechanics and its Applications 206, 553 (1994).DOI
  4. [4]V. Yukalov and S. Gluzman, Self-Similar Bootstrap of Divergent Series, Physical Review E 55, 6552 (1997).DOI
  5. [5]V. Yukalov and S. Gluzman, Self-Similar Exponential Approximants, Physical Review E 58, 1359 (1998).DOI
  6. [6]V. Yukalov and E. Yukalova, Self-Similar Structures and Fractal Transforms in Approximation Theory, Chaos, Solitons & Fractals 14, 839 (2002).DOI
  7. [7]Z. Bai, Lecture notes on determinant quantum monte carlo (dqmc), University of California, Davis, Department of Computer Science (2009), accessed: 2023-10-27.
  8. [8]E. LOH and J. GUBERNATIS, Chapter 4 - stable numerical simulations of models of interacting electrons in condensed-matter physics, in Electronic Phase Transitions, Modern Problems in Condensed Matter Sciences, Vol. 32, edited by W. HANKE and Y. KOPAEV (Elsevier, 1992) pp. 177–235.

Paper details

vlnc:W_sjrbnvhk
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