⚛️ lattice

Diffusion Models for Sampling Near Criticality in Lattice Field Theories

This paper demonstrates that generative diffusion models, trained on small lattice volumes and validated against hybrid Monte Carlo methods, can effectively sample near-critical lattice $\phi^4$ theories across various phases and generalize to unseen larger volumes without retraining, offering a scalable solution for large-volume sampling in lattice field theories.

Original authors: Yang-yang Tan, Gert Aarts, Diaa E. Habibi, Biagio Lucini, Lingxiao Wang

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang-yang Tan, Gert Aarts, Diaa E. Habibi, Biagio Lucini, Lingxiao Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake the perfect loaf of bread, but the dough is so sticky and complex that every time you try to mix it, it gets stuck to the bowl. In the world of physics, this "sticky dough" is a mathematical model called lattice $\phi^4$ theory, which scientists use to understand how particles behave and how the universe changes during phase transitions (like water turning to ice).

For decades, the standard way to mix this dough was a method called Hybrid Monte Carlo (HMC). Think of HMC as a very careful, slow-handed baker who stirs the dough one tiny spoonful at a time. The problem? As the dough gets closer to a critical moment (like the exact temperature where ice melts), it gets incredibly sticky. The baker has to stir millions of times just to get the dough to move a little bit. This is called "critical slowing down," and it makes simulations incredibly slow and expensive.

The New Idea: A "Denoising" Chef

In this paper, the authors (Yang-yang Tan, Gert Aarts, and their team) tried something different. Instead of stirring slowly, they used a Diffusion Model.

Think of a diffusion model like a chef who knows how to un-mess up a photo.

The Forward Process: Imagine taking a clear, perfect photo of your dough and slowly adding static noise to it until it's just a blurry, gray mess.
The Reverse Process: The AI learns to look at that blurry mess and guess, "If I remove a little bit of noise here, and a little bit there, what did the original photo look like?"
The Result: By starting with pure random noise (the blurry mess) and letting the AI "denoise" it step-by-step, the model generates a brand new, perfect loaf of dough instantly, without needing to stir the sticky bowl millions of times.

The Big Test: Does the AI Actually Know the Recipe?

The authors didn't just hope the AI worked; they put it through a rigorous taste test. They compared the AI's "loaves" against the gold standard: the slow, careful HMC baker.

What they found:

The Good News: In both 2D and 3D versions of the theory, the AI successfully recreated the main features of the dough. It got the "magnetization" (how much the dough wants to point in one direction) and the "action density" (the energy of the dough) mostly right.
The Catch: The AI wasn't perfect. In the 3D "broken phase" (where the dough has two distinct flavors), the AI's loaves were slightly too "fluffy" in the middle. This showed up as a small error in the susceptibility (a measure of how easily the dough changes shape). The AI overestimated this by about 59% in one specific 3D test, meaning the peaks of the dough were a bit too wide compared to the real thing.
The Action Density Shift: In 3D, the AI also consistently made the dough slightly too energetic, overestimating the energy density by about 19% to 6% depending on the conditions.

The Magic Trick: Learning from Small Batches to Bake Big Loaves

The most exciting part of the paper is what they call Cross-L Generalization.

Usually, if you train a chef to bake a small 4-inch cake, they can't suddenly bake a giant 64-inch cake without retraining. But the authors trained their AI on a mix of small lattices (sizes 4, 8, 16, and 32) and then asked it to bake a giant 64 lattice it had never seen before.

The Result:

The AI didn't just guess; it actually worked. The model trained on the small sizes successfully generated the giant L=64 lattice.
In fact, in some cases, the "multi-size" trained AI did a better job than an AI trained specifically on the giant size alone.
Why? The authors suggest that by seeing many different sizes, the AI learned the "rules of the dough" (the local physics) better. It learned that the rules don't change just because the pan gets bigger. It saw the "zero-mode" (the overall shape of the dough) fluctuate in small batches, which helped it understand how to handle the giant batch without getting confused.

What They Ruled Out

The paper is very careful about what this technology isn't.

It's not a magic wand: The AI doesn't solve the problem perfectly. It still has errors, especially in the "zero-mode" (the big, slow-moving parts of the system) and the action density in 3D.
It's not just about the biggest training size: They tested if the AI just learned from the largest size it saw (32) and guessed the rest. They found that training on multiple sizes (4, 8, 16, 32) was crucial. A model trained only on size 32 failed to generalize as well as the one trained on all sizes.
It's not a finished product: The authors explicitly state that while the AI generates independent samples (no autocorrelation), it still has a "residual bias." It's not a perfect replacement for the slow baker yet; it's a very fast, very good approximation that sometimes needs a little "tuning" (like a Metropolis correction) to be exact.

How Sure Are They?

The authors are confident in their simulations. They ran thousands of tests, compared the AI's output to the gold-standard HMC data, and measured the errors precisely.

They demonstrated via simulation that the AI can reproduce the "propagator" (how particles talk to each other across the lattice) from the smallest to the largest distances.
They measured that the AI's "acceptance rate" (a test of how well it understands the physics) is high (around 80-90% in good conditions), meaning the AI's guesses are physically sound.
They observed that the "effective sample size" (a measure of how many useful samples the AI produces) is very high, meaning the AI isn't just churning out garbage; it's producing high-quality data.

The Bottom Line

This paper shows that Diffusion Models are a viable, powerful tool for simulating complex physics. They can learn the rules of the universe from small, cheap simulations and apply them to huge, expensive ones without needing to retrain. While they aren't perfect yet (they still get the "fluffiness" of the dough slightly wrong in 3D), they offer a promising shortcut to solving problems that have been too slow to tackle for decades. The authors suggest this could eventually help simulate even more complex things, like full Quantum Chromodynamics (QCD), by training on small lattices and scaling up.

Technical Summary: Diffusion Models for Sampling Near Criticality in Lattice Field Theories

Problem Statement
Lattice field theory (LFT) requires sampling field configurations from the Boltzmann distribution $p(\phi) \propto e^{-S[\phi]}$ to compute non-perturbative expectation values. Standard Markov Chain Monte Carlo (MCMC) methods, such as Hybrid Monte Carlo (HMC) and Metropolis updates, suffer from "critical slowing down" near continuous phase transitions. As the correlation length $\xi$ diverges, the integrated autocorrelation time scales as $\tau_{\text{int}} \sim \xi^z$ (with $z > 1$ ), rendering simulations of large volumes computationally prohibitive. While cluster algorithms and Fourier acceleration mitigate this for specific systems (e.g., $Z_2$ models), a general-purpose solution for interacting lattice field theories across phases and lattice sizes remains elusive. The authors investigate whether generative diffusion models can serve as efficient, independent samplers that bypass Markov-chain autocorrelations, particularly in the challenging near-critical regime.

Methodology
The study focuses on the real scalar $\phi^4$ theory on hypercubic lattices in two and three dimensions, a system exhibiting a second-order phase transition in the Ising universality class.

Model Architecture: The authors employ variance-exploding (VE) score-based diffusion models. A neural network $s_\theta(\phi, t)$ is trained to approximate the score function $\nabla_\phi \log p_t(\phi)$ along a forward noising process defined by a stochastic differential equation (SDE). The forward process adds Gaussian noise to the data distribution until it becomes a broad, nearly Gaussian distribution.
Training Objective: The network is trained via Denoising Score Matching (DSM) to minimize the difference between the learned score and the analytical score of the conditional noised distribution. The architecture is a fully convolutional U-Net with circular padding to enforce periodic boundary conditions, allowing weights to be shared across different lattice volumes ( $V = L^d$ ).
Sampling: After training, samples are generated by integrating the reverse-time SDE (or a probability-flow ODE) starting from Gaussian noise. The authors utilize an Euler–Maruyama discretization for the reverse SDE.
Validation and Diagnostics:
- Reference: Generated ensembles are validated against reference ensembles produced by Fourier-accelerated HMC combined with Wolff cluster updates (FA-HMC–Wolff).
- Observables: Comparison includes scalar observables (magnetization, susceptibility, Binder cumulant), the joint distribution of magnetization and action density, and the momentum-space propagator $G(|k|)$ .
- Local Diagnostics: Two novel diagnostics are introduced: (i) a direct comparison of the learned score against the negative action gradient $-\nabla_\phi S$ on clean configurations, and (ii) the Metropolis-adjusted Langevin algorithm (MALA) acceptance rate using the learned score as a proposal drift.
- Efficiency Metric: An HMC-referenced effective sample size (ESS) is defined based on the mean-squared error (MSE) of observables, combining bias and variance to quantify the quality of the generated ensemble relative to independent target samples.

Key Results

In-Distribution Performance:
- 2D ( $L=128$ ): The diffusion model reproduces scalar observables and the momentum-space propagator across symmetric, near-critical, and broken phases with high fidelity. Residual biases are minimal, concentrated primarily in the zero-mode susceptibility in the broken phase (where the model slightly overestimates the intra-sector width).
- 3D ( $L=64$ ): The model reproduces the propagator and most scalar observables. However, a systematic bias is observed in the action density (overestimated by $\sim 6-19\%$ ) and the zero-mode susceptibility in the broken and near-critical phases. This bias stems from a few-percent error in single-site moments which is amplified by cancellations in the action at $\lambda=0.9$ .
- Convergence: The learned score accurately captures the direction of the action gradient early in training, while the magnitude converges more slowly. The MALA acceptance rate for the converged models exceeds 85% in 2D and $\sim 90\%$ in 3D, indicating the learned drift is a high-quality proposal for exact local updates.
Cross-Volume Generalization:
- 2D: A model trained on $L \in \{8, 16, 32, 64\}$ successfully samples the unseen $L=128$ lattice. In the symmetric and broken phases, it matches or slightly improves upon a model trained directly on $L=128$ . The main degradation occurs at the near-critical point regarding the zero-mode susceptibility.
- 3D: A model trained on $L \in \{4, 8, 16, 32\}$ is applied to the unseen $L=64$ lattice. It reproduces the propagator from ultraviolet to infrared momenta across the phase diagram. Crucially, at the near-critical point, the cross-volume model outperforms the in-distribution $L=64$ model for several zero-mode and single-site observables.
- Mechanism: The authors interpret this success as the multi-volume training exposing the network to finite-volume infrared fluctuations (including the zero-mode distribution) at various scales. This constrains the long-wavelength part of the score more effectively than training on a single large lattice size. Single-size extrapolation baselines confirm that the advantage is not merely due to the largest training size but the combination of multiple scales.

Significance and Claims
The paper establishes that generative diffusion models can serve as viable samplers for lattice field theories near criticality, offering a potential route to bypass critical slowing down. The primary contributions are:

Validation of Diffusion Models: The study demonstrates that reverse-SDE samplers can reproduce complex field-theoretic observables, including the momentum-space propagator and joint distributions, with residual errors largely confined to the zero-mode and specific action density components in 3D.
Cross-Volume Generalization: The work provides quantitative evidence that a fully convolutional score network trained on small lattices can generalize to significantly larger, unseen volumes. This suggests that expensive training on large reference lattices may not be strictly necessary; instead, scores learned from "cheap" small-lattice configurations can transfer to target volumes.
Diagnostic Framework: The introduction of local diagnostics (score gradient alignment, MALA acceptance) and an HMC-referenced ESS provides a rigorous toolkit for assessing the quality of learned samplers without always requiring a reference ensemble for every metric.

The authors conclude that while the method is not yet a complete solution for all observables (notably the 3D action density and zero-mode susceptibility in specific regimes), the cross-volume generalization capability represents a significant step toward scalable sampling for large-volume lattice theories, including potential future applications to full QCD.