Design and optimization of near-field thermophotovoltaic systems using deep learning

Extensive research has been conducted on near-field radiative heat transfer (NFRHT) due to its wide range of applications in energy conversion, radiative cooling, and thermal diodes. The main objective of studying NFRHT at the nanoscale gap is to enhance system performance. This research proposes a new approach to designing and optimizing a near-field thermophotovoltaic (NFTPV) system using deep-learning techniques. Our study utilizes a fully connected network (FCN) and an automated-machine-learning (AutoML) model to simulate radiative heat transfer, aiming to improve radiative heat flux, power generation, and overall system efficiency. By comparing two emitter configurations, we find that the hyperbolic emitter outperforms other configurations, as evidenced by its impact on various system-performance parameters. Significant achievements have been made through our investigations. For the $\mathrm{Si}\mathrm{C}$-plate emitter configuration, we have achieved a notable power density of $0.589\phantom{\rule{0.2em}{0ex}}\mathrm{W}/{\mathrm{cm}}^{2}$ and an efficiency of 23% after accounting for nonradiative recombination at a temperature difference of 600 K with a 100-nm gap. Moreover, the four-period emitter configuration has yielded even more impressive results, with a power density of $0.9452\phantom{\rule{0.2em}{0ex}}\mathrm{W}/{\mathrm{cm}}^{2}$ and an efficiency of 30% after accounting for nonradiative recombination. This study demonstrates the immense potential of utilizing FCN and AutoML for theoretical modeling and optimization of structural parameters in NFTPV systems, as well as providing an accurate model for predicting photocurrent generation. By highlighting the capabilities of these advanced techniques, we have hopefully paved the way for further advancements and innovations in NFRHT.