Generative marginal learning for finite-size molecular simulation with application to stress calculation

We propose generative marginal learning (GML), a framework for equilibrium sampling and local observable estimation in molecular simulations. By training a conditional diffusion model on large-system simulations, GML learns the marginal distribution of a retained region, enabling configuration generation without explicitly simulating the surrounding environment or imposing periodic boundary conditions on the retained region. We derive a marginal effective potential and the associated conditional mean-force identity, which provide the basis for both denoising and force-alignment training. As an application, we estimate the spatially averaged Irving--Kirkwood stress near a vacancy defect. Numerical results demonstrate that GML substantially improves pair structure and local stress accuracy compared with small periodic molecular dynamics simulations, mitigating finite-size effects in calculating local molecular observables.