The inverse design of photonic surfaces produced by high-throughput femtosecond laser processing is limited by a strongly non-linear, many-to-one mapping from laser parameters (power, speed, hatch spacing) to the resulting optical spectrum. Tandem Neural Networks (TNNs) address this challenge by training separate forward and inverse models, but require artificial noise to find diverse solutions, and still provide limited exploration of the design space. We propose Direct Gradient Optimization (DGO), a single-network alternative that treats the pre-trained forward surrogate as a differentiable proxy for the laser-material interaction and back-propagates errors to the process parameters. Two optimization modes are assessed: single-start DGO and Tournament-DGO, which launches multiple random seeds, runs a brief qualification phase, and refines only the five most promising candidates. Across 10,000+ inverse design tasks, Tournament-DGO cuts the average spectral root mean squared error (RMSE) from 1.29% (best TNN) to 0.70%, and boosts design novelty (NEPD) from 0.26 to 0.38.
Haboub et al. (Mon,) studied this question.