Thermal emission modulation of fabrication-friendly, free-form metasurfaces via explainable deep-learning Bayesian optimization

Abstract

Free-form metasurfaces with superimposed transformative meta-atoms provide a versatile platform to realize cross-band thermal emission control. However, design and manufacturing of free-form metasurfaces is extremely challenging, owing to the complex and fractal sub-wavelength topology. Here, we address these two issues by proposing an explainable deep-learning Bayesian optimization (DeepBO) framework to realize a library of fabrication-friendly, free-form metasurfaces with different light–matter interaction bandwidths. The DeepBO requires only 50 training data and is capable of screening high-dimensional design space of 1043 thermal photonic structure candidates with bandwidths from 0.3 to 3.2 eV. We unfold the black-box of deep-learning process by pattern recognition and identify the sub-space key features in the high-dimensional design space, which provides insights for thermal photonic metasurface design. We showcase the design and manufacturing of the broadband solar absorber and the narrowband thermophotovoltaic emitter with record-high spectral efficiency. The spectral selectivity of the fabricated free-form metasurface matches well with the design. The fabrication-friendly, free-form metasurfaces realized in this work can be generalized to thermal emitters for broad-ranges applications in energy and sensing.