Paper Overview
- Field: Machine Learning
- Authors: Majdi Hassan, Marta Skreta, Dominique Beaini, Kirill Neklyudov
- Published: 2026-11-09
- arXiv: 2611.12880
- Paper link: https://arxiv.org/abs/2611.12880
Summary
Stochastic differential equations (SDEs) describe stochastic processes across physics, biology, finance, and many other scientific domains. While single-step numerical integration is cheap, observing meaningful events or transitions may require millions of sequential integration steps, resulting in enormous wall-clock simulation times and poor hardware utilization.
Deep learning offers an alternative: amortizing many integration steps into a single neural network forward pass, potentially speeding up simulation dramatically. However, existing supervised methods typically require long simulated trajectory data, which limits their applicability.
This work proposes Kolmogorov-Chapman Maps (KC-Maps) — a method for learning the transition kernel of an SDE that relies only on trajectory-free marginal samples together with the known drift and diffusion coefficients. Given an initial time and state, KC-Maps model the distribution of the process integrated to any future time, i.e., the weak solution of the SDE.
To recover the correct dynamics, KC-Maps satisfy two key properties:
1. Boundary conditions — embedded directly into the neural network architecture. 2. The Kolmogorov-Chapman equation — which provides an efficient training objective.
Experiments show that KC-Maps learn to integrate the dynamics of peptide molecular systems, outperforming classical molecular dynamics simulation in terms of GPU-hours.
Original Abstract
> We introduce Kolmogorov-Chapman Maps (KC-Maps), a method for learning the transition kernel of an SDE that relies only on trajectory-free marginal samples together with the known drift and diffusion coefficients, enabling much faster molecular dynamics simulation without requiring long simulated trajectory data.