Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems
Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems
Authors: Sicheng Ma, Tianyue Yang, Xiuzhe Wu, Xiao Xue Date: 2026-05-07 Paper ID: openalex:2605.05975
Summary
This paper addresses the computational latency of high-fidelity flow reconstruction in scientific machine learning by proposing a consistency-distilled flow-matching approach. By distilling an optimal-transport teacher into a one-step student model, the authors achieve significant speedups and memory efficiency while maintaining high reconstruction fidelity. The method allows for conditional inference by initializing trajectories from low-fidelity observations, demonstrating robustness on fluid benchmarks including Smoke Buoyancy, Turbulent Channel Flow, and Kolmogorov Flow.
Key Contributions
- Introduces a consistency distillation framework that converts optimal-transport flow-matching models into efficient one-step generators for high-fidelity flow reconstruction.
- Demonstrates a 12x inference speedup and superior SSIM performance (23.1% improvement) compared to directly-trained one-step consistency models on complex fluid dynamics benchmarks.
- Enables conditional reconstruction from low-fidelity observations by initializing generative trajectories from noised inputs, without requiring retraining of the teacher model.
Open Questions & Future Work
Key Concepts
- consistency-distilled-flow-matching: A distillation framework that compresses optimal-transport flow-matching teacher models into one-step consistency models for efficient scientific flow reconstruction.
Archivist Review
I approved the consistency-distilled flow-matching concept as it provides a generalizable distillation paradigm for low-latency scientific reconstruction. The open question regarding adaptive inference noise schedules was also approved as it addresses a fundamental limitation in the robustness of current one-step generative approaches in physical settings. Fluid benchmarks mentioned were rejected as they represent domain-specific test cases rather than standardized foundation datasets.
Approved Concepts
- Consistency-Distilled Flow Matching: This represents the primary methodology for enabling high-fidelity, one-step reconstruction for dynamical systems, which is the core contribution of the paper.
Approved Open Questions
- Adaptive Inference Noise Schedules: The current reliance on manually tuned, fixed inference parameters limits the robustness of one-step reconstruction models across diverse scientific datasets. Improving the adaptive scheduling of inference-time noise injection is technically critical to maintaining physical fidelity without compromising the speed advantages of consistency models.
Rejected Candidates
- [dataset] Smoke Buoyancy (
smoke-buoyancy) - low_impact: This is a specific fluid benchmark, not a widely used, critical dataset deserving a permanent vault note. - [dataset] Turbulent Channel Flow (
turbulent-channel-flow) - low_impact: This is a specific fluid benchmark, not a widely used, critical dataset deserving a permanent vault note. - [dataset] Kolmogorov Flow (
kolmogorov-flow) - low_impact: This is a specific fluid benchmark, not a widely used, critical dataset deserving a permanent vault note.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.05975
- paper_id
- 2605.05975
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-augmented-generationlatent-distribution-matching-ldm
- architectures
-
- datasets
-
- concept_slugs
- consistency-distilled-flow-matching
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-10T06:09:08Z