README.md

metadata

library_name: pytorch
tags:
  - pytorch
  - photonics
  - electromagnetics
  - surrogate-model
  - scientific-computing

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

from neon import Neon
model = Neon.from_pretrained()
result = model.predict(
    thickness=0.30, epsilon_real=2.25, wavelength=0.80
)
print(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