Solver Hand-off Frequency and Placement in Hybrid Neural–Physics Cloth Simulation

We study how a hybrid cloth simulator should share work between a fast graph-network emulator and a slower implicit mass–spring solver. Holding the amount of physics fixed, we vary how often control switches between the two solvers and where the physics frames are placed, across three scenarios (one unseen in training) and three training seeds. Switching has no measurable cost, but the placement of physics frames matters substantially; spreading them evenly matches or beats the best tuned single block without needing to know where it should go, and detector-gated routing performs worse than random placement. At this resolution the hybrid is not faster than the reference simulator, so this is an accuracy study of hand-off timing.