Recursive Flow Matching

Generative models have emerged as a powerful paradigm for solving physics systems and modeling complex spatiotemporal dynamics. However, achieving high physical accuracy without incurring high computational cost remains a fundamental challenge, as existing approaches face a critical speed-fidelity trade-off. In this work, we introduce Recursive Flow Matching (RecFM), a generative framework for forecasting complex spatiotemporal dynamics. RecFM enforces self-consistency to align trajectories across discretization scales, reducing discretization errors and improving performance across metrics for physics-based tasks. To our knowledge, this is the first method to achieve high-fidelity one- and few-step (2-4 step) dynamic generation for scientific systems with performance comparable to state-of-the-art multi-step solvers. Across challenging scientific benchmarks, RecFM achieves up to a 20times speedup over leading diffusion-based emulators while improving predictive accuracy. Furthermore, RecFM reduces mean squared error by over 15% compared to vanilla flow matching, offering a scalable and efficient solution for real-time scientific emulation.

📋 한 줄 요약

**[Generative Spatiotemporal / Flow Matching]** RecFM이 self-consistency로 trajectory를 discretization scale에 걸쳐 정렬해 1-step·2-4 step 생성 가능 — leading diffusion 대비 최대 20× speedup·vanilla flow matching 대비 MSE 15%+ 감소, scientific emulation의 실시간 솔루션.

🎯 핵심 기여도

생성 모델이 물리 시스템·복잡 spatiotemporal dynamics 모델링의 강력 패러다임으로 부상했지만, high computational cost 없이 high physical accuracy 달성이 fundamental challenge — speed-fidelity trade-off에 직면한다고 진단.
RecFM(Recursive Flow Matching) 도입 — 복잡 spatiotemporal dynamics forecasting의 생성 프레임워크.
Self-consistency를 enforce해 trajectory를 discretization scale에 걸쳐 정렬, discretization 오차 감소·physics 기반 task의 메트릭 전반 성능 개선.
저자들이 아는 한 첫 번째로 scientific system에 대해 high-fidelity 1-step·few-step(2-4 step) dynamic 생성을 SOTA multi-step solver와 comparable 성능으로 달성.

💡 핵심 아이디어

Scientific dynamics 생성에서 speed-fidelity trade-off는 flow matching trajectory를 discretization scale 전반에 걸쳐 self-consistent하게 정렬하면 깨질 수 있으며, 이를 통해 1-step·few-step 생성이 SOTA multi-step solver와 comparable한 fidelity를 달성하면서도 substantial speedup을 얻는다.

🔬 기술적 접근법

**방법론**: RecFM — Recursive Flow Matching with self-consistency.
**핵심 기법**: (1) Flow matching trajectory에 self-consistency 강제, (2) 여러 discretization scale에 걸친 trajectory 정렬, (3) Discretization 오차 감소로 physics 기반 정확도 향상, (4) 1-step·few-step(2-4 step) high-fidelity 생성, (5) Multi-step solver와 comparable 성능.

📊 주요 결과

Scientific 벤치마크에서 leading diffusion 기반 emulator 대비 최대 20× speedup, 동시 predictive accuracy 개선.
Vanilla flow matching 대비 MSE 15%+ 감소.
1-step·2-4 step 생성으로 SOTA multi-step solver와 comparable.
실시간 scientific emulation을 위한 scalable·efficient 솔루션 제시.

💭 의의 및 한계

**의의**: Scientific dynamics의 generative emulation에서 speed-fidelity dilemma 해소, self-consistency라는 simple yet powerful 정렬 원리 정립, 첫 1-step few-step high-fidelity scientific 생성, 실시간 emulation 가능. **한계**: 검증된 scientific benchmark 범위 명시 부족, self-consistency loss의 hyperparameter 민감도, 극히 복잡한 multi-physics 시스템 확장 검증 필요.

🚀 실용적 활용

실시간 유체·기상·플라즈마 시뮬레이션.
과학 데이터 augmentation·forecasting.
Diffusion 모델의 step compression 일반 기법.