AI estimation of partition functions with the PULSE method for the thermodynamic properties of disordered compounds

arXiv:2605.28594 · cond-mat.stat-mech, cs.AI, physics.comp-ph · Submitted 2026-05-27 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "AI estimation of partition functions with the PULSE method for the thermodynamic properties of disordered compounds".

Jane: The gist The PULSE method is an improved generative tool that estimates thermodynamic properties of chemically disordered compounds by sampling and estimating the partition function,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, we're wrapping up this look at the PULSE method for estimating partition functions in disordered compounds—this paper is basically about using machine learning to cut through some really messy math when studying materials where atoms are randomly placed <ref:2605.28594#pg1>.

Jane: It’s by the authors who have been working on unsupervised learning sampling and evaluation, calling it PULSE, and they’re showing how this technique can be generalized beyond just simple local atomic quantities to any observable property you want to measure <ref:2605.28594#pg2>.

Lu: What this means is that researchers aren't stuck calculating one specific energy or defect size anymore; they can estimate the whole thermodynamic behavior of the material using this AI model <ref:2605.28594#pg1>.

Meng: The authors prove it works by testing it on the two-dimensional Ising model, which is a benchmark system that has an exact solution, so they have a perfect yardstick to check against their AI predictions <ref:2605.28594#pg1>.

Lalam: And the main result is that PULSE gets within one percent error of the exact value by using only two hundred to two thousand samples per temperature, which is a massive reduction compared to traditional Monte Carlo algorithms <ref:2605.28594#pg1>.

Tom: That’s a big win for computational efficiency because it means you can get reliable thermodynamic predictions with significantly less simulation time than before, even in those tricky high-temperature ranges where they perform best <ref:2605.28594#pg1>.

Jane: It really changes the game for studying complex disordered systems because it moves us away from needing an astronomical number of configurations to understand bulk behavior <ref:2605.28594#pg1>.

Lu: But they did have to mention a caveat, which is important for anyone planning to use this; they found that low-temperature estimates are slightly less accurate because the error builds up across the different partition functions used to compute those final results <ref:2605.28594#pg1>.

Meng: From a production standpoint, that means we need to keep an eye on how this model handles those low-temperature regimes if we want it to be ready for real-world deployment, because error accumulation can be a headache <ref:2605.28594#pg1>.

Lalam: I'm pretty sure this work shows PULSE is a solid step forward in getting accurate estimates for chemically disordered compounds using way less computational effort than the standard methods we use now <ref:2605.28594#pg1>.

Tom: So, it’s about using generative AI to intelligently sample and estimate properties for materials that were previously too hard to simulate accurately <ref:2605.28594#pg1>.

Jane: It’s about shifting the focus from brute-force searching the configuration space to an AI method that gets you what you need with much less computational cost <ref:2605.28594#pg1>.

Lu: The next thing we should look at is how they plan to incorporate prior physical knowledge into this model, because right now it’s purely data-driven and might struggle with very specific new cases <ref:2605.28594#pg1>.

Meng: Yeah, I want to know what kind of training data they suggest would be most useful if we wanted to push this toward industrial use, because just being accurate on a benchmark doesn't mean it runs smoothly in production <ref:2605.28594#pg1>.

Lalam: I’m pretty sure this work marks a step forward in addressing the properties of chemically disordered compounds with a relatively low number of samples <ref:2605.28594#pg1>.

Conclusion: Tom: So, we're wrapping up this look at the PULSE method for estimating partition functions in disordered compounds—this paper is basically about using an AI method called PULSE to estimate partition functions for materials that have disorder, which is basically just atoms randomly swapped around in a crystal lattice.

Meng: So the main idea is that this AI model can generate representative configurations while giving you a direct estimate of how the system behaves thermodynamically, which saves you from running millions of simulations.

Lalam: They tested it on the two-dimensional Ising model, which is a perfect benchmark because it has an exact solution and shares the same math as real disordered materials.

Tom: And what they found is that PULSE gets pretty accurate results, under one percent error compared to the exact value, using only two hundred to two thousand samples per temperature.

Jane: That’s a big deal because it means you can get reliable thermodynamic predictions with way fewer simulations than before, which is what we really want for studying complex materials.

Lu: They also showed that the method scales well, meaning they can take bigger systems, like a sixty by sixty square lattice, and still get convergence without needing more samples relative to the size of the configuration space.

Meng: That scalability is good for production because it means the cost of running this AI estimation doesn't explode just because you need a bigger material model.

Lalam: It’s definitely a significant step forward in estimating properties for these chemically disordered compounds with much lower computational effort than before, which is what I’m pretty sure they are confident about.

Tom: The main thing here is that this generative approach provides a way to study disordered materials at scales that were previously too hard to reach with conventional methods.

Jane: So, it’s about moving from brute-force searching the configuration space to an AI method that intelligently samples and estimates what we actually need for real science.

Lu: The next thing to watch is how they handle those low-temperature results, because the authors flag that those estimates accumulate errors across different partition functions, making them slightly less accurate than the high-temperature ones.

Meng: I’m curious about what breaks at scale; does this AI approach need a specific type of training data to be useful for real industrial applications?

CEA DES IRESNE DEC Cadarache F-13108 Saint-Paul-Lez-Durance France · Universite Paris-Saclay CEA LIST F-91120 Palaiseau France

cond-mat.stat-mech, cs.AI, physics.comp-ph

Submitted: 2026-05-27

Updated: 2026-10-19

Comments: 13 pages, 11 figures, submitted to Computational Materials Science

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The gist The PULSE method is an improved generative tool that estimates thermodynamic properties of chemically disordered compounds by sampling and estimating the partition function, which reduces

Key concepts

Chemical Disorder: This refers to the exponential increase in possible atomic arrangements in a crystal lattice due to the random placement of different atomic species. This complexity makes calculating material properties extremely difficult using standard methods.
Partition Function: This is a central quantity in statistical mechanics that summarizes all possible states of a system at a given temperature. Estimating it accurately allows researchers to calculate crucial thermodynamic properties like energy and susceptibility.
PULSE Method: PULSE is an unsupervised machine learning tool based on Variational Autoencoders (VAE). It learns the underlying physics by generating statistically representative atomic configurations while simultaneously providing estimates for the partition function, making it efficient.

Terminology

Adaptability — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Unsupervised Learning — 1× in this paper · explained in 2 episodes

Summary

The gist The PULSE method is an improved generative tool that estimates thermodynamic properties of chemically disordered compounds by sampling and estimating the partition function, which reduces computational cost compared to traditional Monte Carlo approaches.

Introduction and Motivation

The aim is to reduce the computational cost of Monte Carlo approaches for this type of material and to demonstrate that this generative tool can estimate thermodynamic properties by sampling and estimating the partition function of the system Chemical disorder refers to the exponential growth in the number of possible configurations of atomic species on the crystal lattice owing to the random occupation of atomic sites Classic Monte Carlo methods, although capable of estimating target properties, are computationally inefficient as they require the generation of a large number of configurations before convergence The introduction of Machine Learning (ML) methods has made it possible to optimize overly complex calculations thanks to their generalization capabilities

PULSE Methodology

The PULSE method is based on unsupervised VAE ML models PULSE training proceeds via an unsupervised loop with no prior dataset The goal of the present study is to improve the PULSE method by generalizing the approach to any partition function defined with respect to an observable property of the system To define PO(x) as the probability for the property to have the value O(x) in the configuration space for a given temperature T and configuration x The learning involves solving an optimization problem based on maximizing a lower bound of the partition function ZO

Ising Model Benchmark

The 2D Ising model appears to be an ideal testing ground for our target approach The system consists of a square lattice of size N = L × L, where each site is occupied by a spin The total energy E of the system is described by the expression E = −J Σ X i,j SiSj − H Σ X N i=1 Si The central interest in the study of this model lies in the analysis of physical quantities such as susceptibility χ and specific heat CV

Results and Comparison

PULSE reaches a remarkable accuracy, with an error below 1% compared to the exact value, using between 200 and 2 500 samples per temperature The performance of the method is therefore satisfactory: accurate predictions of the target properties can be obtained with few samples and thus low computational cost In Fig. 6, we demonstrate the scalability of our method by extending it to larger Ising system sizes (from 3×3 to 60×60) In light of these results, we conclude that PULSE is robust with respect to system size: without changing the size of the model, very few samples (relative to the size of the configuration space) are needed to achieve convergence PULSE provides a significantly better estimate of the canonical partition function of the Ising system, while requiring fewer samples than Wolff Most notably, it is usually the hightemperature range, where PULSE performs best, that is the most interesting for many applications including nuclear ones

Conclusion

PULSE represents a significant advance in addressing the properties of chemically disordered compounds With a relatively low number of samples, PULSE provides excellent estimates of the target property, outperforming the capabilities of the benchmark Monte Carlo methods in the high-temperature regime Low-temperature estimates are slightly less accurate due to accumulation of errors on the various partition functions combined to compute the target property Future developments could include incorporating prior knowledge about the physical environment to improve the model performance in handling very specific cases

How it works

The method learns physics through its partition function
Training proceeds via an unsupervised loop where configuration energies are computed and used to determine error for backpropagation
The learning involves solving an optimization problem based on maximizing a lower bound of the partition function ZO

Improvements for AI systems

Improved estimation of thermodynamic properties for chemically disordered systems using a generative approach PULSE method. This allows AI systems to estimate observable properties of an atomic-scale physical system formulated as partition functions, such as thermodynamic properties for materials like high-entropy alloys or (U,Pu)O2 MOX fuels.
Enhanced computational efficiency for sampling disordered materials by minimizing required configurations. The paper demonstrates that the method can estimate properties by sampling and estimating the partition function of the system, achieving this with a relatively low number of samples compared to traditional Monte Carlo methods.
Scalable performance across different system sizes for Ising model simulations. The improved PULSE method is robust with respect to system size, showing that the amount of configurations needed in the sampling procedure does not exceed 2 000 samples per temperature, and convergence improves as the number of spins increases.
Accurate prediction of specific thermodynamic observables through partition function recombination. The system can estimate properties like modified magnetic susceptibility by combining different partition functions, though it notes that properties that are combination of more than two partition functions, as in Eq. (19), present slight deviations at low temperature.
Differentiated performance based on temperature regimes for various sampling methods. The AI system can be tuned to perform best in specific regimes: PULSE performs best at high temperatures, while other Monte Carlo methods like Wolff's algorithm are better at low temperatures (e.g., Wolff’s algorithm performs better at low temperatures).