Traditional time integrators struggle with nonsmooth systems, often slowing down or losing accuracy. Our new approach uses Reinforcement Learning to automatically choose optimal time steps, adapting even when the dynamics are nonsmooth. Tested on three challenging problems, a sliding-mode control system, a diode circuit, and a seismic fault model, our RL-based integrator significant speed-ups, reaching up to 10× faster in the hardest cases. A promising new direction for efficient time integration across complex dynamical systems.
For more details: https://link.springer.com/article/10.1007/s44379-025-00048-6