Machine Learning and Bayesian Optimization in the Analysis of Van der Waals Bilayers

ABSTRACT Van der Waals homobilayers, comprising two identical layers of a 2D material, are significant due to their tunable electronic properties and potential applications in advanced electronic and optoelectronic devices, such as transistors, photodetectors, and solar cells. This study utilizes machine learning (ML) and optimization methods to predict the bandgaps of these homobilayers. The Extra Tree Regressor (ETR), which features 48 attributes, provides reasonable accuracy. To enhance predictive performance, we integrate three optimization approaches evaluated across 1000 experiments. The ETR model achieves precise results while necessitating a reduced number of density functional theory (DFT) calculations. The Gaussian Process (GP) model, in conjunction with Bayesian optimization, improves prediction by utilizing probabilistic outputs and uncertainty assessments. We subsequently employed a hybrid methodology that combines ETR and GP models: initially utilizing ETR for fast exploration and then shifting to GP for enhanced optimization. The suggested hybrid method exhibits efficiency and accuracy, attaining robust prediction ability while utilizing fewer computer resources ‐ showing its applicability in future research.