Machine Learning Enhanced Design Optimization and Knowledge Discovery for Multi-Junction Photonic Power Converters

We compare some classical and machine-learning enhanced design optimization methodologies. We investigate the design of the complex structures of ten-junction InP latticematched photonic power converters with In0.53 Ga0.47 As absorbers optimized for operation at 1550 nm with 53.6% ± 1.3% conversion efficiency. We find that the implicit pattern recognition capabilities of dimensionality reduction using principal component analysis accelerates design discovery, optimization, and the understanding of complex optical phenomena in the simulated devices.