FlowForge: A Staged Local Rollout Engine for Flow Field Prediction

Learning-based surrogate simulation for time-dependent physical flows must be accurate, robust to noisy data, and fast enough for repeated use. Existing surrogates often rely on complex models and update the entire field in one pass, which can be slow and fragile to imperfect data. We introduce FlowForge, a staged local rollout surrogate that predicts future flow fields by orchestrating local updates in a carefully designed sequence. This design allows FlowForge to simulate flow fields using lightweight models while maintaining strong accuracy. Across incompressible, compressible, and multiphase benchmarks from CFDBench, PDEBench, and BubbleML, FlowForge matches or improves upon strong baseline families, while reducing per-step latency and showing consistently higher robustness to noise and missing observations.