A machine learning-quantum chemical quest to predict the photovoltaic properties of 2D molybdenum sulfide super structures

The current study introduces a machine learning (ML)-driven method to forecast the photovoltaic (PV) characteristics of two-dimensional (2D) molybdenum sulfide (MoS) superstructures. Structural data of about 1202 compounds is compiled and four top performing models are trained to predict their open-circuit voltage (V_oc). The models are evaluated by using their mean squared errors (MSE), which spanned from 0.47 to 1.23, reflecting adequate predictive accuracy. Their stability evaluations are determined via their tolerance factor which reveals a significant relationship with both V_oc. Moreover, the materials with its optical properties are confirmed using the Kramers-Kronig relations, displaying a high level of consistency between experimental data and simulated results. The SHapley Additive exPlanations (SHAP) analysis identifies that their polarizability (Alpha) and first order connectivity (chi1) are the most important factors to affect their model performance. The bandgaps of MoS₂ and MoS₃₃ are calculated as 1.47 and 0.72 eV, respectively. These results can enhance the understanding of the PV potential of MoS based 2D materials and can lay the foundation for further studies in this field.