Action Learning for Scientific Simulations
Field: Machine Learning Authors: Sarah Liaw, Yuanqi Du, Yilun Du Published: 2026-10-22 arXiv: 2610.31585
Abstract
Understanding and controlling how systems evolve is central to science and engineering. Existing approaches learn local dynamics or generate trajectories, but objectives from individual transitions can miss dependencies across time, while generative models do not provide a tractable objective for optimizing arbitrary paths. We study action learning: recovering a stochastic action from trajectory observations, yielding a common objective for reasoning about dynamical systems. We introduce LEAP (Learned Energy on Action Paths), which anchors the unknown action to a tractable reference process and learns the remaining correction by comparing observed trajectories with reference trajectories. Once learned, the same action can be flexibly reused for path ranking, mechanism steering, constrained optimization, and sampling.
Summary
- Problem: Objectives built from single transitions may miss dependencies across time; generative trajectory models lack a tractable objective for optimizing arbitrary paths.
- Approach: LEAP (Learned Energy on Action Paths) anchors the unknown stochastic action to a tractable reference process, then learns the residual correction by comparing observed trajectories with reference trajectories.
- Key property: The result is a differentiable objective over path space, capturing both local dynamics and non-additive cross-time dependencies.
- Reusability: Once learned, the same action supports path ranking, mechanism steering, constrained optimization, and sampling.
- Experiments: On stochastic physical systems, many-body orbital dynamics, and molecular simulations, LEAP recovers underlying dynamics and enables transfer across objectives and conditions.