Inverse thin-film design

The network proposes. Physics decides.

Ink a reflectance curve, or pick one. Sixty-four random eight-layer TiO₂/SiO₂ coatings descend through a small neural network toward it — the neural-adjoint method — and then an exact transfer-matrix solver re-scores every one. The winner is chosen by the physics, not the network.

or try:

Stack · 8 layers

Looks like

Verdict

descending…

Sixty-four guesses at an eight-layer stack, poured through the network at once. Watch them settle toward the green band — then I check every one against the exact solver.

An honest note: for eight thin layers the exact solver takes microseconds, so the network isn't needed for speed here. It stands in for the expensive solvers — RCWA, FDTD — where surrogates earn their keep. Because the exact answer is instant in this toy, every guess the network makes can be audited live, including the ones it gets wrong.

Simplifications: lossless layers, normal incidence, no reflection from the back of the glass. The neural-adjoint method is from Ren et al.⁠ (opens in new tab) (NeurIPS 2020); I applied it to metasurface inverse design in Neural-adjoint method for the inverse design of all-dielectric metasurfaces (Optics Express 2021).