Abstract
In lattice field theory, Monte Carlo simulation algorithms get highly affected by critical slowing down in the critical region, where autocorrelation time increases rapidly. Hence the cost of generation of lattice configurations near the critical region increases sharply. In this paper, we use a Conditional Generative Adversarial Network (C-GAN) for sampling lattice configurations. We train the C-GAN on the dataset consisting of Hybrid Monte Carlo (HMC) samples in regions away from the critical region, i.e., in the regions where the HMC simulation cost is not so high. Then we use the trained C-GAN model to generate independent samples in the critical region. We perform both interpolation and extrapolation to the critical region. Thus, the overall computational cost is reduced. We test our approach for Gross-Neveu model in 1+1 dimension. We find that the observable distributions obtained from the proposed C-GAN model match with those obtained from HMC simulations, while circumventing the problem of critical slowing down.
Introduction: The Challenge of Critical Slowing Down
Lattice field theory is a cornerstone for simulating fundamental forces, yet it faces a significant computational hurdle: critical slowing down. When using traditional Hybrid Monte Carlo (HMC) algorithms to sample field configurations near a phase transition (the "critical point"), the autocorrelation time—the number of steps required to produce an independent sample—spikes exponentially. In some regimes, the cost scales as $a^{-z}$ with $z > 6$, where $a$ is the lattice spacing. This makes high-precision simulations near the continuum limit extremely expensive.
This research explores using machine learning, specifically Conditional Generative Adversarial Networks (C-GANs), to bypass this bottleneck. By training on "cheap" data from non-critical regions, the model aims to interpolate or extrapolate into the expensive critical regions, generating independent samples nearly instantaneously.
Methodology: Physics-Informed C-GANs
The authors utilize a C-GAN architecture conditioned on the theory's coupling constant, $\lambda$. The study uses the Gross-Neveu (GN) model in 1+1 dimensions on a $32 \times 32$ lattice as a testbed. The GN model is a popular "toy model" for Quantum Chromodynamics (QCD) because it shares features like asymptotic freedom and chiral phase transitions.
To ensure the model captures the underlying physics, several key innovations were implemented:
- Symmetry Preservation: Convolutional layers and periodic padding were used to respect the translation symmetry and topology of the lattice.
- Data Pre-processing: A natural log transformation, $\sigma'_i(x) = \ln(\sigma_i(x) + c)$, was applied to the auxiliary field $\sigma$ to stabilize training and manage the dynamical range.
- Training Strategy: The model was trained on 10 ensembles of 4,000 HMC samples each, taken at $\lambda$ values where HMC is efficient (e.g., 0.6, 0.8, 2.2, 2.5). This avoids the critical region ($\lambda \approx 1.5$) during training to test interpolation capabilities.
Key Results and Insights
The C-GAN approach successfully addressed the critical slowing down problem by providing zero autocorrelation by design. Since GANs generate samples from random noise, each output is statistically independent.
Observable Matching and Phase Transitions
The model's performance was validated by comparing physical observables against baseline HMC results. The C-GAN accurately reproduced the Chiral Condensate ($|\bar{\sigma}|$) and Susceptibility ($\chi$) across the entire range of $\lambda$, including the critical region.
Notably, the susceptibility peak, which characterizes the second-order phase transition in the GN model, was correctly captured even when the model had to interpolate into the critical region.
The Importance of Domain Knowledge
Ablation studies revealed that "vanilla" ML models are often insufficient for scientific tasks. Without periodic padding and the log-transformation, the GAN failed to reproduce correct physical observables, highlighting the necessity of physics-specific "guardrails" in the architecture.
Conclusion and Future Outlook
This work demonstrates that C-GANs can effectively "fill in the gaps" in lattice simulations, generating high-quality configurations at a fraction of the traditional computational cost. While successful in 1+1 dimensional toy models, the next challenge is scaling this approach to 3+1 dimensional Lattice QCD, which involves complex gauge fields and more demanding symmetries. Future research will also investigate the model's ability to handle the "sign problem" and more complex fermionic actions.
Summary updated