Machine Learning Framework for Predicting Strain-Induced Electronic and Optoelectronic Properties of Heterostructure TMDCs

The electronic and optoelectronic properties of heterostructure Transition Metal Dichalcogenides (TMDCs) are highly sensitive to external perturbations such as strain, which offers a promising route for tuning material functionalities. However, performing density functional theory (DFT) calculations, especially with GW-BSE level accuracy, for a wide range of strain conditions and heterostructure configurations is computationally expensive and time-consuming, creating a bottleneck in the discovery of new materials. In this work, we present a machine learning (ML) framework capable of predicting strain-induced modifications to the band structure and excitonic properties of TMDC heterostructures with high accuracy. Our approach leverages a dataset generated from high-throughput density functional theory calculations, incorporating variations in strain, stacking configurations, interlayer distance and surface chalcogenide vacancies.

We utilize a neural network architecture optimized to predict key electronic features such as bandgap, exciton binding energy, and bandgap type (direct/indirect) under different strain conditions. The model will significantly reduce computational costs compared to traditional DFT approaches of GW-BSE calculations. Additionally, we explore the model's interpretability to gain insights into the underlying physics governing strain-dependent changes in electronic and optoelectronic properties. This framework not only accelerates the discovery of tunable TMDC-based materials for next-generation optoelectronic devices but also provides a blueprint for surrogate ML models for traditional ab initio methods.