Recursive Flow Matching: Generative AI for Physics Simulations
A new generative AI framework called Recursive Flow Matching (RecFM) achieves high-fidelity forecasting of complex spatiotemporal dynamics with up to 20x speedup over leading diffusion-based emulators. Introduced in arXiv preprint 2605.26535, RecFM enforces self-consistency across discretization scales to reduce errors, enabling one- and few-step (2-4 step) generation comparable to multi-step solvers. This is the first method to balance speed and accuracy for physics-based tasks, addressing the critical speed-fidelity trade-off in existing approaches.
Key facts
- Recursive Flow Matching (RecFM) is a generative framework for forecasting complex spatiotemporal dynamics.
- It enforces self-consistency to align trajectories across discretization scales.
- RecFM achieves up to a 20× speedup over leading diffusion-based emulators.
- It is the first method to achieve high-fidelity one- and few-step (2-4 step) dynamic generation for scientific systems.
- Performance is comparable to state-of-the-art multi-step solvers.
- The work is published on arXiv with ID 2605.26535.
- It addresses the speed-fidelity trade-off in generative models for physics systems.
- The framework reduces discretization errors and improves performance across metrics.
Entities
Institutions
- arXiv