A unified evaluation framework for benchmarking scientific machine learning surrogates (neural operators and Transformer-based foundation models) that predict steady-state incompressible Navier-Stokes flow around complex 2D geometries. Its core contribution is a single score, normalized to 0-100 on a logarithmic MSE scale, that integrates three complementary metrics: global mean-squared error, near-boundary MSE (boundary-layer fidelity), and PDE residual (physical consistency), so heterogeneous models and geometric input representations (binary occupancy mask vs. signed distance field) can be compared on one axis. Benchmarking on the FlowBench lid-driven-cavity dataset shows that foundation models generally outperform neural operators, SDF inputs help given sufficient training data, and all models degrade under geometry and Reynolds-number distribution shift.
This paper benchmarks scientific machine learning (SciML) models for predicting steady-state incompressible Navier-Stokes fluid flow (x/y velocity and pressure fields) around complex two-dimensional …