Thermophotovoltaic emitter design with a hyper-heuristic custom optimizer enabled by deep learning surrogates

Abstract

Micro/nanostructures hold promise for use as emitters to boost the efficiency of thermophotovoltaic (TPV) systems. For example, periodic gratings may alter the spectrum of irradiation on the photovoltaic cell to better match the spectral response of the cell. Photons with energies slightly higher than the bandgap of the semiconductor are the most desired as they generate electron-hole pairs with minimal thermalization losses. This prompts the use of gratings as selective emitters, and the choice of grating geometry has been an active research topic. Even for a one-dimensional (1D) grating, millions of possible geometries exist, and each grating requires computationally intensive full-wave simulation, e.g., the rigorous coupled-wave analysis (RCWA), to calculate the spectral, directional emissivity. Since optimization algorithm performance is problem-dependent, in this work, a hyper-heuristic search enabled by a fully connected neural network surrogate of the native RCWA is employed to select an algorithm for a specific grating design problem. A comparison with existing untuned algorithms for an ideal emitter problem demonstrates that the hyper-heuristically generated algorithm yields superior performance. This algorithm is then employed for the optimization of the emitter for a full TPV system comprising a heated 1D tungsten binary grating paired with a 300 K InGaSb cell. The system is optimized for maximum power or efficiency at 2000 K and 1500 K, respectively, and the grating properties for the optimized cases are analyzed.