Skip to content

ARTICLE SUMMARY / 2026

Physics-constrained neural networks for surrogate modeling of lossless periodic structures

E. Prehn & P. Jung

arXiv · 2606.28119v2 · 10 September

Summary

Can a learned RCWA surrogate respect an exact physical constraint?

Projects predicted scattering outputs onto a manifold that enforces energy conservation.

Demonstrates differentiable inverse design for a diffractive waveguide combiner.

Commentary

The stated model is lossless; its constraint does not transfer unchanged to absorbers or amplifiers.

Physical consistency is valuable, but does not guarantee accuracy for unfamiliar geometry or unresolved resonances.

Interpretation and connections in this note, separate from the authors’ reported findings.

Related modelling approaches

Relation labels come from the method guide and are editorial. Unlabelled links are conceptual connections or comparison methods; the paper does not necessarily implement them.

CONNECTIONS

Related notes

READ ALONGSIDE IT

Related reading notes

2026 / Numerical method

Matrix Square Root Based Differentiable RCWA Implementation for High-Performance Parallel Computing

F. Van der Ceelen, Y. Shao & W. Coene

Replaces eigendecomposition with matrix-square-root and matrix-exponential evaluations.

Summary & commentary : Matrix Square Root Based Differentiable RCWA Implementation for High-Performance Parallel Computing