The research question
Can a learned model predict optical fields well enough to be useful in photonic engineering? Our experiments so far work under specific conditions. The model does not replace classical solvers in general.
Conditions belong with the result
Boundary information
Our best results are given the reference field on the boundary of each tile. We state that condition with every comparison.
Generalization
Performance on one setup does not establish reliable prediction from scratch on a different geometry.
Domain size
Larger stitched domains remain a research question rather than an assumed capability.
Practical speed
Any speed claim must include preparation, inference and comparison conditions; a practical advantage remains unresolved.
What a useful comparison needs
A fair experiment states its geometry, inputs and boundary information, reference solution, error measure and full runtime conditions, and it shows the misses as well as the successes. We will publish numbers here once they meet that standard.
The next research questions
Next we investigate prediction with less boundary information, behaviour across different geometries, larger domains and an end-to-end speed comparison. These are research questions, not scheduled features.
Examples & validation