Manufacturability-Constrained Surrogate Inverse Design of Disordered Photonic Crystals for Ultrathin Silicon Photovoltaics

Ultrathin crystalline-silicon films use far less silicon than conventional wafers, but a film hundreds of nanometers thick transmits much of the sunlight a wafer would absorb. Patterning the film with a deliberately disordered photonic crystal recovers much of that loss, and we show that the specific arrangement of holes is itself a design variable: layouts of identical disorder strength differ by up to 1.6\% in absorbed solar flux. From 2,723 simulated layouts (each satisfying an etchable 50~nm minimum wall and a fixed fill fraction), we train an ensemble of convolutional neural networks to predict a layout's absorption enhancement directly from its geometry, with 0.21\% mean error at over four orders of magnitude lower cost per evaluation. A constrained search uses the network to propose and the full solver to verify, and counts a gain only above a measured precision floor; it beats an equal-budget random search in all eight disorder settings tested, and the largest verified gain reaches 2.6\% above its setting's dataset average. Every design satisfies the fabrication constraints throughout, and the method could extend to other patterned-film design problems. Code and the Photra-2.7k dataset are available at https://anonymous.4open.science/r/SEER-Photonic-Design/.