no. 16 · Tuesday 17 November 2026
Structure-Preserving Neural Time-Stepping for Gradient-Flow PDEs via Onsager's Variational Principle
Abstract
We propose a unified Onsager-based neural time-stepping (ONTS) framework for solving gradient-flow partial differential equations with conserved and nonconserved dynamics. Unlike conventional neural time-stepping methods that directly parameterize the evolving state, the proposed framework employs spatial neural networks to approximate the Onsager process variables, namely, the rates in nonconserved systems and the fluxes in conserved systems. Within this Eulerian framework, we develop three complementary schemes: an explicit least-squares scheme (E-ONTS), an implicit variational scheme (I-ONTS), and a predictor–corrector scheme (PC-ONTS) that combines explicit prediction with safeguarded implicit correction. For conserved systems, the flux-based formulation guarantees discrete mass conservation under periodic or no-flux boundary conditions. Furthermore, we rigorously establish unconditional discrete energy stability for the implicit schemes, provided that a suitable descent criterion is satisfied, without requiring global minimization of the neural variational objective. Numerical experiments for the heat, Allen–Cahn, Cahn–Hilliard, Fokker–Planck, and coupled Poisson–Nernst–Planck equations demonstrate the accuracy, robustness, and structure-preserving properties of the proposed framework.
- comments
- 35 pages, 8 figures
- fields
- Numerical Analysis