Machine-learning-assisted metasurface design for high-efficiency thermal emitter optimization

Nanophotonic devices can provide solutions to challenges in energy conversion, information technologies, chemical or biological sensing, quantum computing, and secure communications. The realization of practical optical structures and devices is a complex problem due to the multitude of constraints on their optical performance, materials, scalability, and experimental tolerances, all of which are requirements implying large optimization spaces. However, despite the complexity of the process, to date, almost all nanophotonic structures are designed either intuitively or based on a priori selected topologies, and by adjusting a limited number of parameters. These intuition-based models are limited to ad hoc needs and have narrow applicability and predictive power, with the exhaustive parameter searches often performed manually. Since the comprehensive search in hyper-dimensional design space is highly resource-heavy, multi-objective optimization has so far been almost impossible. Humans' restrained capacity to think hyper-dimensionally also limits the perception of multivariate optimization models, and, therefore, advanced machinery is needed to manage the multi-domain, hyper-dimensional design parameter space. In this work, we merge the topology optimization method with deep learning algorithms, such as adversarial autoencoders, and show substantial improvement of the optimization process in terms of computational time (4900 times faster) and final devices efficiencies (∼98%) by providing unparalleled control of the compact design space representations. By enabling efficient, global optimization searches within complex landscapes, the proposed compact hyperparametric representations could become crucial for multi-constrained problems. The proposed approach could enable a much broader scope of the optimal designs and data-driven materials synthesis that goes beyond photonic and optoelectronic applications.