AI and the World's Hard Problems · Chemistry & materials · arXiv · 2026-10-23
Synthetic Stochastic Time Series from Generative Models
Obtaining statistically converged time-series data is costly in computational chemistry, because high-fidelity simulations are expensive. Four generative architectures were compared, and a denoising diffusion model reproduced the target statistics most consistently and generated faster than a sequential Transformer. For donor-acceptor energy-gap trajectories, augmenting 0.1 ns of training data gave accuracy comparable to 0.25 to 0.50 ns of direct molecular dynamics sampling.