Gradient-informed Bayesian and interior point optimization for efficient inverse design in nanophotonics

Inverse design provides a systematic approach for developing high-performance nanophotonic devices. Although numerous optimization algorithms exist, previous global approaches exhibit slow convergence, and conversely, local search strategies frequently become trapped in local optima. To address the limitations inherent to both, we introduce BONNI: Bayesian optimization through neural network ensemble surrogates with interior point optimization. It augments global optimization by efficiently incorporating gradient information to determine optimal sampling points. We demonstrate BONNI's capabilities in the design of a distributed Bragg reflector and dual-layer grating coupler through a comparison with other commonly used optimization algorithms.