Neon Slab Predictor
A fast neural network predictor for dielectric slab optical response, trained on FDFD simulation data from the Neon solver.
What it does
Give it a dielectric slab configuration. Get optical response back in milliseconds.
Inputs: slab thickness (um), relative permittivity real part, wavelength (um)
Outputs: normalized transmission, normalized reflection, normalized peak intensity
Install and use
1from neon import Neon
2model = Neon.from_pretrained()
3result = model.predict(
4 thickness=0.30, epsilon_real=2.25, wavelength=0.80
5)
6print(result)
What this is
This model is trained on 2D scalar FDFD simulation data from the Neon benchmark solver. It covers one geometry class: a rectangular dielectric slab at normal incidence in vacuum. Within that class it is a fast, usable predictor.
What this is not
This model does not generalize to metasurfaces, waveguides, photonic crystals, multilayer stacks, oblique incidence, dispersive materials, or full-vector Maxwell problems. If your structure is not a simple dielectric slab, this model will give you a number that means nothing.
Performance
Saved single-model test MAE:
- transmission: 0.053586
- reflection: 0.055240
- intensity: 0.234785
Saved OOD degradation warning:
- benchmark-facing Model C ensemble OOD transmission MAE: 0.107476
- benchmark-facing Model C ensemble OOD reflection MAE: 0.105841
- benchmark-facing Model C ensemble OOD intensity MAE: 0.174486
- overall OOD mean MAE: 0.129268
Note: the repository does not currently store a single-model benchmark-facing OOD summary. The OOD warning above comes from the saved 5-member Model C ensemble evaluation.
Training data range
- thickness: 0.12 to 0.46 um
- epsilon_real: 1.4 to 4.0
- wavelength: 0.72 to 0.92 um
- inputs outside this range trigger validation;
predict() raises a validation error by default and can warn instead with warn_only=True
Companion paper
Toward Trustworthy Surrogate Models for Electromagnetic Simulation: A Systematic Evaluation of Physics-Informed Training, Uncertainty, and Active Learning on a Controlled Benchmark
in preparation
Citation
TBD
