Selective Thermophotovoltaic Emitter with Aperiodic Multilayer Structures Designed by Machine Learning
Wenbin Zhang, Boxiang Wang, Changying Zhao
Abstract
Controlling the thermal emission characteristics of a thermophotovoltaic (TPV) emitter can effectively reduce the thermal losses in a photovoltaic cell. Moreover, selective TPV emitters can facilitate the generation of electricity from a variety of high-temperature waste heat sources. To overcome the challenge of achieving perfect spectral selectivity of TPV emitters, the development of artificial intelligence provides a vision for the optimization of selective TPV emitters beyond the conventional paradigm. In this work, we demonstrate that a highly selective, aperiodic TPV emitter with a high figure of merit (FOM) can be achieved with the help of Bayesian optimization. The design of the selective TPV emitter is optimized over 5.23 × 109 candidate structures in multilayers consisting of multiple components to maximize the FOM. The maximum FOM could be realized within calculations for less than 0.67% of the total candidate structures, which is much better than other machine learning algorithms. As for the gallium antimonide photovoltaic cell, the resulting optimal structure is an aperiodic multilayer structure with an FOM of 82.16%. The optimal structure is then fabricated by a multi-target sputtering system, and the experimental result of the emission characteristics is achieved with an FOM of 81.35%, which is significantly better than those of multilayers with similar material designed and fabricated based on previous studies. What is more, we have analyzed the efficiency of the TPV system and measured the thermal stability of the fabricated samples. The results demonstrate that Bayesian optimization is efficient in designing selective TPV emitters, and the machine-learning-based design of metamaterials can be extended for the expensive black-box global optimization problems in other field applications.