Generative Sampling of Reaction Intermediates using Physics Regularized Autoencoders

Understanding reaction intermediates and transitions in complex molecular systems remains a fundamental challenge in statistical physics and computational chemistry. In this work, we explore the use of physics-regularized autoencoders as a generative tool for sampling such intermediates from high-dimensional dynamical data. We explore several autoencoder architectures with a variety of loss functions on synthetic trajectories generated via Brownian and Langevin dynamics. These test systems are designed to exhibit key physical phenomena such as entropic switching, where the dominant reaction pathway shifts due to thermal effects. By incorporating physical constraints and priors into the training process, we assess each model's ability to recover the relevant dynamics and generate physically consistent intermediate states. Results are benchmarked against numerical simulations of the corresponding backward Kolmogorov equations of the system.