Modulating electronic properties of semiconductor materials at large mechanical deformation

Elastic strain engineering explores the full 6D space of admissible ultra-large strains and the effects on physical properties. However, the complexity of controllably engineering materials properties by mechanical forces necessitates first-principles computations to design an optimal straining pathway. In our work, to map the 6D strain space, we developed a general machine learning framework that adopts convolutional neural networks, physics informed data representation scheme, and a new active learning algorithm to allow bandgap and band structure prediction, band extrema detection, and effective mass calculations for semiconductor materials. Combining this method with experimentally validated finite-element simulations, we identified the most energy-efficient strain pathways that would reversibly transform an ultrawide-bandgap material to a metalized state without phonon stability. The fast and reliable inference of the proposed model opens a path towards analyzing and scrutinizing general band structures in the vast 6D strain space.