PhysicsNeMo Challenger Track
A cross-branch benchmark program for learned Physics-AI representations. PhysicsNeMo is evaluated against the POD reference within each parameter branch rather than becoming a separate branch itself.
PhysicsNeMo is a challenger to the representation/model layer, not a parameter branch
The parameter branches answer “what operating space are we learning?” PhysicsNeMo answers “what learned representation/model family should be benchmarked against the interpretable POD reference?” Keeping those axes separate makes the roadmap much easier to reason about.
POD family
Interpretable, low-dimensional and already validated on the Speed Sweep. It remains the control model even after learned challengers are introduced.
PhysicsNeMo
Modern learned-representation framework to be tested on exactly the same accepted CFD, holdouts and engineering metrics.
Trust / UQ / OOD
Sits above both model families. A stronger predictor is not automatically a more trustworthy engineering system.
Where PhysicsNeMo enters the NACA0012 program
What must remain fixed when PhysicsNeMo results are added
Data and holdouts
- same accepted CFD cases
- same preprocessing/provenance contract
- same LOSO or held-out parameter conditions
- same interpolation/extrapolation labels
Fields, forces and deployment
- velocity + pressure/surface quantities
- Cd/Cl derived consistently
- oracle/representation gap where meaningful
- latency/resource cost and trust behavior