Abstract
In recent years, the use of end-to-end neural networks to analyze Monte Carlo data has received a lot of attention. However, the application of non-end-to-end generative adversarial neural networks is less explored. Here, we study classical many-body systems using generative adversarial neural networks. We use the conditional generative adversarial network with an auxiliary classifier (AC-GAN) and introduce self-attention layers into the generator, enabling the model to learn the distribution of two-dimensional XY model spin configurations as well as the physical quantities. By applying the symmetry of the systems, we further find that AC-GAN can be trained with a very small raw dataset, allowing us to obtain reliable measurements in the model that requires a large sample size, e.g. the large-sized 2D XY model and the 3D Heisenberg model. We also find that it is possible to quantify the distribution changes that occur in the configuration of the models during phase transitions and locate the phase transition points by AC-GAN.
Introduction
In condensed matter physics, Markov Chain Monte Carlo (MCMC) simulations are the standard for studying many-body systems. However, these simulations face significant challenges, such as "critical slowdown" near phase transitions and high computational costs for large-scale 2D or 3D systems. While deep learning has emerged as a potential solution for generating system configurations, standard convolutional neural networks (CNNs) often fail to capture the long-range spatial correlations essential for accurate physical modeling. This paper explores the use of an Auxiliary Classifier Generative Adversarial Network (AC-GAN) enhanced with self-attention to overcome these limitations.
Methodology: AC-GAN and Self-Attention
The authors implement an AC-GAN architecture where the Generator (G) produces spin configurations based on random noise and a target temperature, while the Discriminator (D) evaluates the authenticity of the configuration and predicts its temperature class.
A critical innovation is the integration of multi-head self-attention layers within the Generator. Unlike standard CNNs, which have a limited receptive field, self-attention allows the model to capture global relationships between spins across the entire lattice. This is particularly vital for modeling non-local physical phenomena like vorticity in the 2D XY model.
Data Efficiency via Physical Symmetry
To minimize the need for expensive MCMC sampling, the study leverages the inherent symmetries of physical systems (translation, reflection, and rotation). By applying these operations to a very small set of raw MCMC samples, the authors can artificially expand the training data. For the 3D Heisenberg model, this approach allowed the AC-GAN to be trained using only 31 raw configurations, reducing the data collection time by over 99% while maintaining high accuracy in physical observables.
Results and Physical Accuracy
The AC-GAN demonstrates remarkable fidelity in reproducing physical quantities such as mean energy, magnetization, and vorticity. It significantly outperforms previous generative models, particularly at low temperatures where long-range correlations are dominant.
As shown in the visual comparison, the AC-GAN successfully captures the transition from ordered states at low temperatures ($T=0.05$) to highly disordered states at high temperatures ($T=2.0$).
Detecting Phase Transitions
The authors propose a novel method for identifying phase transitions by analyzing the distribution overlap (termed "CrossArea") between MCMC and GAN-generated samples. GANs often exhibit a form of "symmetry breaking" at low temperatures, where they generate configurations aligned in a single direction, unlike the more varied MCMC samples. By measuring the gradient of this distribution discrepancy, the authors can accurately pinpoint critical temperatures ($T_c$). For the 2D XY model, this method yielded $T_c \approx 0.9$, closely matching the theoretical value of 0.89.
Conclusion
This work demonstrates that AC-GANs, when augmented with self-attention and physical symmetry constraints, are powerful tools for simulating many-body systems. The ability to train on extremely sparse datasets and the introduction of the interpretable "CrossArea" metric for phase transition detection offer promising avenues for studying complex physical systems where traditional MCMC methods are computationally prohibitive.
Summary updated