Shared Physics Responses Recover Hidden Rankings in Neural Operator Libraries

Choosing the best neural‑operator model when you lack a high‑fidelity reference solution is a common obstacle in scientific computing. Researchers often resort to expensive simulations or intuition to decide which checkpoint to keep. A new approach shows that, under a squared Hilbert‑space loss, the relative performance of a finite set of operators can be inferred from a much smaller, physics‑based quantity.

What You Need to Know

The paper proves that ranking models in a library depends only on the low‑dimensional subspace spanned by the differences between candidate predictions. Because this subspace is small, a single linearized response of the governing equation—computed once with an anchor model—provides enough information to score every model in the library simultaneously. The authors call this the “shared physical diagnostic.”

They test the method on Fourier‑based and convolution‑based neural operators applied to three classic PDE families: incompressible fluid flow, reaction‑diffusion systems, and wave propagation. Across these test beds, the diagnostic recovers more than 99.6 % of the pairwise preferences that would be obtained by brute‑force comparison and identifies the optimal checkpoint in over 99.0 % of cases.

Why It Matters

In many deployment scenarios, generating a reference solution is prohibitive—either because the underlying physics is too costly to simulate at high resolution or because experimental data are scarce. Being able to trust a cheap, physics‑based proxy means practitioners can select the best model without additional costly runs, saving both time and computational resources.

Moreover, the shared diagnostic sometimes yields a corrected physical proxy that outperforms the best individual candidate in the library. This suggests that the linearized response can capture complementary information from multiple models, effectively acting as an ensemble estimator without the overhead of storing or evaluating many networks.

Key Details

• Uses squared Hilbert‑space loss as the ranking criterion.
• Relies on the anchor‑based linearized response of the governing equation.
• Evaluated on Fourier neural operators and convolutional neural operators.
• Tested on fluid dynamics (Navier‑Stokes), reaction‑diffusion (Gray‑Scott), and wave equations.
• Achieves >99.6 % recovery of pairwise preferences and >99.0 % identification of optimal checkpoints.

What’s Next

Future work could extend the diagnostic to other loss functions, explore adaptive anchor selection, and investigate whether the shared response can be used for online model selection during simulation. Applying the technique to larger, multi‑physics datasets will clarify its limits and potential as a routine tool for neural‑operator deployment.

📌 Source: Arxiv Ml