Bayesian optimization-driven enhancement of the thermoelectric properties of polycrystalline III-V semiconductor thin films

• Published in: NPG Asia materials Vol. 16; no. 1; pp. 17 - 7
• Main Authors: Ishiyama, Takamitsu, Nozawa, Koki, Nishida, Takeshi, Suemasu, Takashi, Toko, Kaoru
• Format: Journal Article
• Language: English
• Published: Tokyo Nature Publishing Group 29.03.2024; Nature Portfolio
• Subjects: Artificial intelligence; Bayesian analysis; Deposition; Dopants; Effectiveness; Figure of merit; Gaussian process; Group III-V semiconductors; Optimization; Parameters; Semiconductors; Substrates; Thermoelectric materials; Thin films
• ISSN: 1884-4057; 1884-4049; 1884-4057
• DOI: 10.1038/s41427-024-00536-w

Abstract Studying the properties of thermoelectric materials needs substantial effort owing to the interplay of the trade-off relationships among the influential parameters. In view of this issue, artificial intelligence has recently been used to investigate and optimize thermoelectric materials. Here, we used Bayesian optimization to improve the thermoelectric properties of multicomponent III–V materials; this domain warrants comprehensive investigation due to the need to simultaneously control multiple parameters. We designated the figure of merit ZT as the objective function to improve and search for a five-dimensional space comprising the composition of InGaAsSb thin films, dopant concentration, and film-deposition temperatures. After six Bayesian optimization cycles, ZT exhibited an approximately threefold improvement compared to its values obtained in the random initial experimental trials. Additional analysis employing Gaussian process regression elucidated that a high In composition and low substrate temperature were particularly effective at increasing ZT . The optimal substrate temperature (205 °C) demonstrated the potential for depositing InGaAsSb thermoelectric thin films onto plastic substrates. These findings not only promote the development of thermoelectric devices based on III–V semiconductors but also highlight the effectiveness of using Bayesian optimization for multicomponent materials.