Machine learning enhanced design and knowledge discovery for multi-junction photonic power converters

Machine learning is proving to be a revolutionary tool across many disciplines, including optoelectronic device design. In this report, we compare classical and machine learning enhanced design optimization methodologies. We investigate, as an example case, the design of the complex structures of ten-junction InP lattice matched photonic power converters with InGaAs absorbers optimized for operation at 1550 nm. We find that the implicit pattern recognition capabilities of dimensionality reduction using principal component analysis accelerate design discovery, optimization, and the understanding of complex optical phenomena in the simulated devices. The dimensionality reduction approach offers over twenty times as many optimal designs with greater variability and with a 15% reduction in computational cost compared to a classical optimization method. Furthermore, we find that the representation of the reduced dimensionality subspace offers an intuitive interpretation of optical phenomena expected to occur in this design problem. This method is general and offers the potential for knowledge discovery, expanded design perspective, and optimization acceleration in conjunction with a significant reduction in computational expense in systems which can be numerically modeled.