Ultrabroadband and band-selective thermal meta-emitters by machine learning

Thermal nanophotonics enables fundamental breakthroughs across technological applications from energy technology to information processing1–11. From thermal emitters to thermophotovoltaics and thermal camouflage, precise spectral engineering has been bottlenecked by trial-and-error approaches. Concurrently, machine learning has demonstrated its powerful capabilities in the design of nanophotonic and meta-materials12–18. However, it remains a considerable challenge to develop a general design methodology for tailoring high-performance nanophotonic emitters with ultrabroadband control and precise band selectivity, as they are constrained by predefined geometries and materials, local optimization traps and traditional algorithms. Here we propose an unconventional machine learning-based paradigm that can design a multitude of ultrabroadband and band-selective thermal meta-emitters by realizing multiparameter optimization with sparse data that encompasses three-dimensional structural complexity and material diversity. Our framework enables dual design capabilities: (1) it automates the inverse design of a vast number of possible metastructure and material combinations for spectral tailoring; (2) it has an unprecedented ability to design various three-dimensional meta-emitters by applying a three-plane modelling method that transcends the limitations of traditional, flat, two-dimensional structures. We present seven proof-of-concept meta-emitters that exhibit superior optical and radiative cooling performance surpassing current state-of-the-art designs. We provide a generalizable framework for fabricating three-dimensional nanophotonic materials, which facilitates global optimization through expanded geometric freedom and dimensionality and a comprehensive materials database.