Home / Research / Sim2Real Physical Fidelity
Accepted · NeurIPS 2026 XAI4Science

Sim2Real Physical Fidelity

Physics-grounded diagnostics for understanding when better neural-operator benchmark scores do not necessarily imply better agreement with physically meaningful flow behavior.

Publication page arXiv
CFD + PIVForecastInterventionPhysical diagnostics

Research question

Can a scientific AI model obtain a better aggregate prediction score while representing important flow statistics less faithfully?

Approach

  • Paired CFD and experimental PIV data around a NACA4418 airfoil.
  • 20 observed PIV frames mapped to 20 future frames.
  • Four evaluated neural-operator architectures: hybrid reference, FNO, CNO, and Transolver.
  • Mean-preserving masking of energetic, random, and low-energy regions with frozen-model re-evaluation.
  • Comparison of field error, two-component fluctuation-energy error, MVPE, regional error, and total fluctuation-energy diagnostics.

Current result

Aggregate prediction scores and physically meaningful flow diagnostics can disagree. The accepted workshop study documents this behavior and defines the limitations that motivate the next stage of the project, including energy-matched controls, smooth interventions, physically realizable solver-based perturbations, and broader cross-system evaluation.

Related publication

Do Better Scores Mean Better Physics? Physics-Grounded Explanations for Sim2Real Neural Operators