Machine Learning for Chemistry and Chemistry for AI
What can artificial intelligence do to accelerate molecular simulations? And what can statistical mechanics tell us about artificial intelligence? In our group, we explore both directions, combining machine learning with ideas from statistical mechanics.
Machine learning for chemistry. Molecular simulations let us follow atoms as they move, but predicting whether a material prefers to be solid or liquid requires understanding both its enthalpy and its entropy. Entropy describes the many ways its atoms can arrange themselves and is notoriously difficult to calculate. We use deep learning to estimate it from short simulations of each phase separately, allowing us to predict melting temperatures without sampling transitions between the phases [1].
Learning when a material melts. Our entropy estimates allow us to compare solid and liquid stability. The temperature where their Gibbs free energy difference crosses zero gives the predicted melting point. Figure 3 from Ref. [1].
We also use generative AI to extract dynamical information from quantum simulations. In collaboration with Daniel Freedman, we developed diffusion models that generate a range of spectra consistent with the simulation data. This lets us estimate uncertainty in properties such as molecular diffusion and identify features that may be artifacts of the reconstruction [2].
Chemistry for AI. Our group also explored what chemistry can do for AI. Training a neural network resembles a search across a complex landscape, much like a molecular system exploring possible configurations. We use first-passage theory, the mathematics of how long a fluctuating process takes to reach a target, to study learning. This approach helps us predict which perturbations can accelerate training and how frequently to apply them [3].
Learning as a first-passage process. The schematic compares training without perturbations (A) and with repeated perturbations (B), showing how we define the time needed to reach a chosen accuracy. Figure 1 from Ref. [3].
These ideas also help us use large language models more efficiently. Our Reset and Discard (ReD) method controls how many attempts a model spends on each problem, redirecting effort when repeated attempts bring diminishing returns. On coding, mathematics, and reasoning benchmarks, it solves more distinct problems within a fixed computational budget [4].
By studying molecules and learning algorithms together, we aim to develop better simulations of matter and a deeper understanding of how models learn.
References
[1] Ben-Shimon, Y., Hirshberg, B. & Bar-Sinai, Y. Machine learning the entropy to estimate free energy differences without sampling transitions. npj Computational Materials 12, 202 (2026).
[2] Meir, S., Freedman, D. & Hirshberg, B. Using Diffusion Models to Estimate Uncertainties in Analytic Continuation. arXiv:2608.13123 (2026), preprint.
[3] Meir, S., Keidar, T. D., Reuveni, S. & Hirshberg, B. First-passage approach to optimizing perturbations for improved training of machine learning models. Machine Learning: Science and Technology 6, 025053 (2025).
[4] Meir, S., Keidar, T. D., Levi, N., Reuveni, S. & Hirshberg, B. More Bang for the Buck: Improving the Inference of Large Language Models at a Fixed Budget using Reset and Discard (ReD). arXiv:2601.21522 (2026), preprint.
