Physical-consistency trade-off of physics-residual distillation

Determine whether physics-residual distillation improves nodal power balance despite increasing prediction error measured by root-mean-square error.

Background

The proposed physics-residual distillation term consistently worsened prediction RMSE in the reported ablation study. However, the experiments did not report direct physics-residual metrics such as mean, 95th-percentile, or maximum nodal residuals. Consequently, the paper leaves unresolved whether the apparently harmful term may nevertheless improve physical consistency in a way that could justify accepting higher predictive error in grid-operation applications.

References

More importantly for a paper whose central term is physical, we report no physics-residual metric, so it remains possible that physics-residual distillation improves nodal power balance while costing RMSE, a trade an operator might accept; adding mean, p95 and maximum nodal residual per configuration would most strengthen the study.

— Physics-Aware Distillation of a Grid Foundation Model for Real-Time Edge Inference  (2610.21833 - Sănduleac et al., 15 Oct 2026) in Section 6, Discussion, paragraph “What these experiments do not establish”