Physics-informed machine learning encodes physical laws as constraints — energy conservation, boundary conditions, symmetry requirements — directly into the model's loss function or architecture. The premise: more physical constraints produce better models. Physics is ground truth; encoding it can only help.
Ray (arXiv:2512.22261) shows that explicit energy conservation enforcement is mathematically redundant when the underlying equations are already physically consistent. The mean error with and without the conservation constraint is approximately 7 × 10⁻⁹ in both cases. The constraint does nothing because the equations already satisfy it. Adding it is adding a tautology.
The sharper finding: Fabry-Perot oscillations — interference effects that are physically real and improve the physical accuracy of simulated optical spectra — degrade machine learning performance. Removing them improves bandwidth prediction R² by 31.3% and reduces error by 73.8%. The physically more complete model is the worse training signal.
The mechanism: Fabry-Perot oscillations introduce high-frequency structure into the spectra that the ML model interprets as features to fit. The model allocates capacity to reproducing the oscillations — which are real physics but irrelevant to the prediction target (bandwidth). The capacity spent fitting oscillations is capacity not spent learning the relationship between structural parameters and bandwidth. The physically accurate features act as a distraction, not a guide.
The structural lesson: physical completeness and ML optimality are different objectives. Physical models aim to reproduce all relevant phenomena. ML models aim to predict specific targets from specific inputs. When the physical model contains phenomena that are real but uncorrelated with the prediction target, those phenomena become noise from the ML perspective — noise that happens to be exactly structured, making it harder to ignore than random noise.
More physics is not always more information. It can be more distraction.
Ray, "The Physics Constraint Paradox: When Removing Explicit Constraints Improves Physics-Informed Data for Machine Learning," arXiv:2512.22261 (2025).