Deep learning assisted spectral control of metasurface antireflection coatings for thermophotovoltaic systems

As a promising technology for sustainable energy development and efficient energy utilization, thermophotovoltaic (TPV) systems have garnered significant attention. This work proposes a multimodal deep neural network (MMDNN)-assisted grey wolf optimization (GWO) framework for computationally efficient inverse design of metasurface antireflection coatings (ARCs), enabling TPV-oriented multi-band spectral tailoring while alleviating the burden of exhaustive simulations. An ARC with micro-nano-pyramid structure is adopted to improve the spectral selectivity of TPV cells by minimizing reflectance within the high-efficiency power generation range (0.68–1.74 μm) while preserving high reflectance in the short-wave band, and a photon-recycling back-surface reflector (BSR) is incorporated to provide high out-of-band reflectance. The electro-optical-thermal multi-physics coupled analysis is conducted on a TPV system incorporating the converter optical stack comprising the front-side ARC and BSR to link spectral modification with coupled carrier and thermal processes. The impacts of structural parameters on ARC reflectance are analyzed to provide a preliminary physical understanding and to define the optimization search space. The MMDNN surrogate is then trained to learn the mapping from structural parameters and wavelength to spectral reflectance, and then embedded in GWO to rapidly identify the optimal geometry. With the optimized front-side ARC and BSR in place under enhanced cooling, the TPV system achieved an output power density of 2943.9 W/m², and the cell's opto-electrical efficiency reached 26.55%. These results highlight the coupled role of spectral control and thermal management in improving TPV energy conversion.