Synthetic Stochastic Time Series from Generative Models

Abstract

Obtaining statistically converged time-series data represents a fundamental challenge in computational chemistry, where sampling is constrained by the prohibitive computational cost of high-fidelity simulations. The accuracy attainable from a finite simulation budget is governed by the standard-error scaling $1/\sqrt{n}$, rendering exhaustive sampling challenging for systems of realistic complexity. Here we show that generative models trained on high-quality stochastic trajectories can reproduce key statistical properties of the underlying process and generate additional synthetic trajectories at substantially lower computational cost, thereby enhancing the sampling accessible from a finite set of simulations. We present a systematic comparison of four generative architectures across two complementary data sources: an analytically tractable Ornstein-Uhlenbeck reference and donor-acceptor energy-gap trajectories from all-atom molecular dynamics simulations. Among the four architectures tested, the denoising diffusion probabilistic model reproduces the target statistics most consistently, across both data sources and across the range of correlation times examined, and generates far faster than the sequential Transformer. Specifically, for the energy-gap trajectories, augmentation from 0.1 ns of training data yields statistical estimates comparable in accuracy to those obtained from 0.25-0.50 ns of direct MD sampling. We further show that the benefit of synthetic augmentation is bounded by the information contained in the training data, producing a quality floor determined by both the training data and the generative model. Our findings establish diffusion-based generative models as an effective route to data-driven augmentation of stochastic time series in computational chemistry.