Small-sample efficient inverse design of anti-reflection micro-nano optical structures: a Bayesian optimization-tuned...

The inverse design of micro-/nano-optical anti-reflection structures is fundamentally limited by the high-dimensional nonlinear parameter mapping and prohibitive computational cost of electromagnetic simulations, while existing data-driven methods suffer from severe small-sample performance degradation. In this work, we propose a closed-loop inverse design framework based on a Bayesian optimization-tuned extremely randomized trees surrogate model for efficient anti-reflection structure optimization. The proposed model achieves an average relative error of only 1.01% for independent test samples, outperforming six mainstream machine learning models. Combined with the differential evolution algorithm, our framework realizes 100% constraint satisfaction for reflectivity ≤20% at 1064 nm across 0°-70° incident angles, while reducing finite element method calls by 99.67% and achieving a ∼307× computational speedup. This work provides an engineering-feasible solution for high-performance optical structure design.