arXiv:2605.07157 · v2

Learned Lagrangian Models of PDEs via Euler-Lagrange Residual Minimization

通过欧拉-拉格朗日残差最小化学习偏微分方程的拉格朗日模型

首次提交 2026-05-08 10:43

版本更新 2026-11-13 02:13

雷达观测 2026-11-13 14:15

摘要

We present the first method to directly use a learned continuous Lagrangian to forecast the dynamics of systems governed by partial differential equations, exploiting the inherent conservative structure to achieve stable long-range predictions. We develop an integrator that minimizes the squared Euler-Lagrange residual via a discretization-agnostic, structure-preserving construction on local space-time patches. Different from integrators for analytical models, integrators for learned models should decouple model error (phase error) from integration error (conservation error). By minimizing the residual locally on overlapping patches, we bypass the global implicit solve inherent to fixed discretizations, which slows time- and space-stepping and complicates learning. Our method scales linearly with domain size via Jacobi iteration, and places no structural requirements on the learned network, allowing it to be coupled with existing physics-guided machine learning (ML) methods. We validate our approach on a learned representation of a double pendulum, a one-dimensional wave equation, and a two-dimensional wave equation. Our method achieves long-horizon predictions with error comparable to classical symplectic methods while generalizing to spatially varying dynamics and arbitrary boundary conditions without retraining.

作者备注：Accepted at NeurIPS 2026 (spotlight). 10 pages, 9 figures, 3 tables, 9 pages of appendices