Sampling the Liquid-Gas Critical Point with Boltzmann Generators Preprint: AIP/123-QED

Luigi de Santis, John Russo, Andrea Ninarello · 2026

Read Sampling the Liquid-Gas Critical Point with Boltzmann Generators Preprint: AIP/123-QED by Luigi de Santis, John Russo, Andrea Ninarello (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-05109.txt.

Introduction

Luigi de Santis and the rest of the team test Boltzmann Generators, generative models sampling balance setups from the Boltzmann spread, where usual runs suffer severe dynamical bottlenecks. Focusing on a Lennard-Jones fluid's liquid-gas key point, they show the generators catch key actions signatures, perform reliably when trained near criticality, and extrapolate across nearby states. The model's leanness metric closely follows underlying phase boundaries, suggesting a link between generative scores and thermodynamics. Small open system sizes remain a limit, suppressing large swings characterizing key phenomena.

Catalog record

CREATORLuigi de Santis, John Russo, Andrea Ninarello
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31712
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.05109

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Sampling the Liquid-Gas Critical Point with Boltzmann Generators Preprint: AIP/123-QED Luigi de Santis, John Russo, Andrea Ninarello Abstract Generative models based on invertible transformations provide a physics-aware route to sample equilibrium configurations directly from the Boltzmann distribution, enabling efficient exploration of complex thermodynamic landscapes. Here, we evaluate their applicability in regions where conventional simulations suffer from severe dynamical bottlenecks, focusing on the liquid–gas critical point of a Lennard–Jones fluid. We show that Boltzmann Generators capture essential signatures of critical behavior, retain reliable performance when trained at or near criticality, and extrapolate across neighboring states of the phase diagram. An intriguing observation is that the model’s efficiency metric closely traces the underlying phase boundaries, hinting at a connection between generative performance and thermodynamics. However, the approach remains limited by the small system sizes currently accessible, which suppress the large fluctuations that characterize critical phenomena. Our results delineate the current capabilities and boundaries of Boltzmann Generators in challenging regions of phase space, while pointing toward future applications in problems dominated by slow dynamics, such as glass formation and nucleation. I Introduction Since the mid-20th century, molecular dynamics and Monte Carlo simulations have been primary tools for exploring the phase behavior of liquids and understanding their equilibrium thermodynamics. 17 ; 31 These techniques remain foundational, as their primary objective is to reconstruct equilibrium distributions by sampling configurational ensembles, whether through stochastic MC moves or deterministic MD trajectories. While simulations excel at exploring accessible regions of phase diagrams, their utility is challenged in systems exhibiting slow equilibration dynamics, metastability, or critical phenomena. Examples include glass-forming liquids hindered by rugged energy landscapes, 8 nucleation processes requiring rare-event sampling, 38 and phase transitions in liquids plagued by critical slowing down. 36 ; 14 ; 40 To address these challenges, enhanced sampling methods such as metadynamics 6 , umbrella sampling, 15 and parallel tempering 22 have emerged that effectively circumvent kinetic barriers by biasing simulations or accelerating phase-space exploration. Yet, the growing complexity of soft matter systems and the demand for high-resolution phase diagrams have spurred interest in integrating machine learning into computational frameworks. Initially, machine learning tools were deployed to classify phases, detect transitions, or analyze order parameters. 39 ; 11 ; 29 However, recent advances have shifted their application toward replacing or enhancing traditional simulation methodologies, particularly in sampling elusive regions of phase space. 28 Notably, ML-driven interatomic potentials now enable accurate and efficient modeling of many-body interactions, 7 while adaptive sampling techniques leverage reinforcement learning to prioritize underrepresented states. 24 Diffusion models, inspired by non-equilibrium thermodynamics, have also shown promise in generating equilibrium configurations by iteratively denoising distributions. 45 Among generative AI approaches, normalizing flows have garnered significant attention. 25 ; 33 Normalizing flows employ sequences of invertible, learnable transformations to map simple base distributions (e.g., Gaussians) to complex target distributions. By embedding thermodynamic principles into their framework, models like Boltzmann Generators (BGs) modify standard normalizing flows to sample directly from the Boltzmann distribution. 32 This is achieved by combining invertible transformations with energy-based learning, allowing the model to generate equilibrium configurations consistent with statistical mechanics. The transformation invertibility ensures exact density estimation, enabling direct sampling of equilibrium states while preserving physical interpretability. Therefore, normalizing flows are able to bridge data-driven learning and physics-informed modeling, offering a transformative framework for exploring challenging regions of phase diagrams. Boltzmann Generators have been recently emerged as powerful tools for sampling complex, high-dimensional distributions in many-particle system models. They have been demonstrated to accelerate MC by combining local moves with learned nonlocal transitions 19 , including in the context of rare-event sampling 4 , enable free energy estimation of atomic solids without requiring ground-truth samples, 44 and facilitate unsupervised training for predicting thermodynamic properties in solids 2 . Extensions to the isobaric-isothermal (NPT) ensemble allow direct sampling of pressure-dependent systems 43 , and their applicability has been demonstrated in studies of glass-forming liquids 21 ; 20 . A recent line of work has shown that Boltzmann Generators can be conditioned on thermodynamic state variables, leading to so-called conditional normalizing flows that enable training within a specific region of a phase diagram and facilitate the generation of configurations across neighboring states 42 ; 34 . Notably, this framework builds on physically informed flow architectures that exploit permutation equivariance to efficiently model transformations between physical distributions. While this approach shows promise for exploring phase space and computing free-energy differences, it remains unclear whether it can effectively address particularly challenging regions, such as those affected by critical slowing down. In this work, we examine this question by employing BGs in the conditional setting, as defined in Ref. 34 , to generate configurations around the liquid-gas critical point of a Lennard-Jones system. We first benchmark the method in a defined liquid phase. We then train BGs directly at the critical point and assess their ability to extrapolate to nearby states. Conversely, we also train away from the critical point and evaluate performance as the system approaches criticality. We validate our findings by examining both energy averages and their distributions. We further compare the expected and generated heat capacity and compressibility, as these fluctuation-based quantities are essential for characterizing critical phenomena. Our results, building on established normalizing-flow methods and architectures, indicate that BGs achieve strong performance close to criticality while still maintaining effectiveness when extrapolated beyond the critical point. However, their efficiency is offset by the small system sizes accessible, since limited access to high-dimensional spaces suppresses large-scale fluctuations. Taken together, these findings delineate the current scope of BGs and point toward their future potential in addressing complex phenomena such as glass formation and nucleation. II Methods II.1 Model We investigate the phase behavior of a Lennard-Jones (LJ) system under isothermal-isobaric (NPT) conditions, focusing on mapping the solid-liquid coexistence line. The system consists of N=180 particles, initially arranged in a disordered, non-overlapping configuration. While the system size might seem limited, it is consistent with current state-of-the-art practices 34 ; 21 . Periodic boundary conditions are applied, and isotropic volume fluctuations are permitted to maintain constant pressure. The LJ potential V_{LJ}=4\epsilon\left[\left(\frac{\sigma}{r}\right)^{12}-\left(\frac{\sigma}{r}\right)^{6}\right] (1) is truncated at a cutoff distance r_{c}=2.2\sigma , with a multiplicative smoothing function, applied between r_{s}=0.9r_{c} and r_{c} to ensure continuity, of the form S(x)=1-6x^{5}+15x^{4}-10x^{3} (2) where x=\frac{r-r_{s}}{r_{c}-r_{s}} . In the following, all results are reported in reduced units of length \sigma and energy \epsilon . In the context of normalizing flows, the prior defines the base distribution from which samples are drawn and subsequently transformed to approximate the target distribution. In our case, the prior state is represented by 20,000 equilibrium configurations obtained through conventional Metropolis Monte Carlo simulations. These configurations were generated over 40 million Monte Carlo steps, with one sample collected every 2,000 steps to ensure statistical independence. In the following, we refer to this distribution as the prior when it is used during training. However, in the conditional setting described below, it may also coincides with the reference distribution against which the network efficiency is evaluated. II.2 Conditional Normalizing Flows As anticipated, a Normalizing Flow is composed of a series of invertible and differentiable transformations, concatenated to transform a simple prior distribution into a target distribution. Interestingly, the prior and target distributions may correspond to either different or

Read the complete text