Physics-informed Bayesian optimization suitable for extrapolation of materials growth

This paper describes a novel physics-informed Bayesian optimization approach that leverages prior physics knowledge, specifically Vegard’s law and the linear relationship between gas flow rate and composition in compound semiconductors. The methodology was applied to metal-organic chemical vapor deposition for III–V semiconductor growth. It resulted in the successful synthesis of III–V semiconductors with tailored band gap wavelengths and lattice constants in the region of growth conditions not included in the training data. Furthermore, it predicted hidden trends that Ga composition would be smaller than In composition in As-rich growth regions. This trend is not described by prior physics, demonstrating that statistical machine learning is effective not only for optimization but also for gaining a physical understanding of crystal growth mechanisms. The study demonstrates the potential to develop extrapolable machine learning models by integrating robust physics knowledge, which significantly enhances the efficiency of high-throughput and autonomous material synthesis.