Herrprofessor/neon-slab-models

Neon: Scalar Slab Benchmark Research Artifact

This repository contains benchmark-specific research artifacts from the Neon codebase. It is not a general-purpose photonics model and it should not be treated as a deployable surrogate.

What is included

The main checkpoint is Model C, a reduced hybrid model trained on Neon’s current 2D scalar normal-incidence rectangular dielectric slab benchmark. A 5-member ensemble and evaluation summaries are also included.

Inputs:

• slab thickness
• slab relative permittivity real part
• wavelength

Outputs:

• benchmark_normalized_transmission
• benchmark_normalized_reflection
• normalized_peak_intensity
• a cropped 145-point complex centerline field

Training data:

• 18 slab designs x 6 wavelengths = 108 total samples
• 72 train, 18 validation, 18 test
• one fixed benchmark geometry, source, grid, monitor layout, and absorber configuration from the Neon slab base case

Intended use

Use these files only to reproduce or inspect the slab-benchmark results reported by Neon.

Limitations

• This model was trained only on the current scalar normal-incidence rectangular dielectric slab benchmark.
• It cannot be used as a general model for metasurfaces, waveguides, photonic crystals, arbitrary multilayer structures, arbitrary source conditions, or full-vector Maxwell problems.
• It does not predict a full 2D field; it predicts slab-response scalars and a cropped centerline field only.
• Any uncertainty values here are ensemble disagreement, not calibrated uncertainty.
• Direct Neon reevaluation remains mandatory for any reported inverse-design candidate.
• The trained weights are a byproduct of an investigation into simulation-driven ML workflows on a controlled benchmark, not a deployable tool.
• The research contribution of Neon is the methodology, validation discipline, and benchmark findings, not the model weights themselves.

Research contribution

Neon’s contribution is the benchmark methodology and the findings about target definition, reduced physics-informed losses, uncertainty diagnostics, active learning, and direct-solver-verified screening on a controlled slab problem.