What a physical constraint buys a learned surrogate

A learned surrogate is attractive when repeated numerical calculations are expensive. It can make exploration practical, but its speed changes how often it is used and how far users may push it. Evaluation therefore needs to describe the operating region, not just an average error on familiar examples from that region.

Physics-informed neural networks provide one way to incorporate governing equations into learning. The original work is a useful reference for thinking about this design choice. A physical constraint supplies information about admissible behavior. It does not, by itself, certify that every prediction is accurate or that the chosen equation fully describes the target system.

The first reading question is where the constraint enters. Is conservation built into the representation, or is a residual penalized during training? These approaches can support different guarantees. A small penalty at selected training points should be interpreted as a measured training outcome, with the sampling locations and weighting kept visible.

Boundary and initial conditions deserve separate attention. A solution can appear plausible in the interior while answering the wrong physical problem at its edges. Testing unfamiliar geometries or operating conditions helps expose that distinction. Such tests should match the intended application rather than serve merely as additional demonstrations on convenient cases.

Long trajectories introduce another issue: local errors can affect later states. Evaluate quantities that matter to the application, including conserved totals or event timing, over the actual duration of use. A single-step comparison cannot describe every consequence of repeatedly feeding a prediction back into the model.

A useful deployment rule couples the surrogate to a fallback. Define conditions that trigger a trusted solver, a diagnostic check, or a refusal to extrapolate. Then account for that fallback when reporting speed. The relevant achievement is a faster scientific workflow with an inspectable error boundary, rather than an isolated inference time.