Neural Time Integrator with Stage Correction

Yifan Zhang, Hong ZhangOpenReview 2025 · Submission to ICLR 2025 · Rejected · OpenReview

Abstract

Numerical simulation of dynamical systems requires time integration solvers that balance accuracy and computational efficiency. Recent work indicates that neural integrators, a hybrid of classical numerical integration and machine learning, can achieve significant performance gains. Building upon this idea, we propose a new type of neural integrator that introduces stage corrections inspired by the fact that traditional time integration schemes such as Runge-Kutta exhibit different error characteristics at each stage. Specifically, our method corrects numerical errors immediately after each stage evaluation by using a neural network, mitigating error propagation across stages. This enables the use of larger time steps while preserving stability and accuracy. We demonstrate that our approach is at least one order of magnitude more accurate than existing hybrid methods for complex nonlinear dynamical systems when integrated with the same step size.
BibTeX
@misc{zhang2025neural,
  author = {Yifan Zhang and Hong Zhang},
  title = {{Neural Time Integrator with Stage Correction}},
  year = {2025},
  url = {https://openreview.net/forum?id=tnSj6FdN8w},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2025},
}
Figure