A new algorithm lets computer simulations decide, on the fly, when to spend on expensive high-fidelity physics and when a cheaper shortcut will do.
Researchers combined three existing techniques: Operator Inference reduced-order models (ROMs), the overlapping Schwarz alternating method for splitting a simulation into subdomains, and reinforcement learning. They trained deep Q-networks offline to choose, subdomain by subdomain, whether to run a full-order model or a pre-trained ROM, balancing accuracy against computational cost and how often the model switches. Once trained, the policy runs on new problems it has never seen, with no need for a reference solution to check against. The team tested it on a 1D advection-diffusion problem with a moving front and a 3D elastic wave propagation problem built in the Norma.jl solid mechanics code.
This matters because most hybrid simulation methods assign a fixed model to each region and leave it there for the whole run. That is a bad fit for transient problems where the interesting physics, a wavefront or a diffusion front, moves through the domain over time. The learned policy tracked that movement, assigning full-order detail to subdomains containing the wave and cheaper ROMs everywhere else, and beat static FOM/ROM assignments on the diffusion benchmark.
It is a promising idea for cutting simulation cost without an engineer manually re-tuning fidelity zones every timestep. But the benchmarks here are still textbook cases, a 1D front and a linear elastic wave, and letting the agent also adapt the domain decomposition itself added no measurable benefit. Whether the approach holds up on messier, nonlinear, real-world engineering models is still an open question.