Data-Driven Framework to Accelerate Semiconductor Discovery

This work presents a Python-based framework for the data-driven acceleration of materials discovery. It aims to autonomously identify regions of compositional space where new materials can be found, which is where more quantitatively accurate modeling techniques can focus. The software integrates an interface to materials databases, machine learning (ML) models based on the composition–property relationship, and data analysis tools. Semiconductor-specific capabilities are included, such as tailored ML feature representations. However, the framework can be applied to any material type. We discuss in detail the implemented ML models and their accuracy in predicting formation energies. These models are then validated on bismuth and lead chalcohalides, where our approach correctly identifies compositional regions where experimentally known materials absent from databases lie. As a side result, we also present thermodynamic convex hulls for Pb–S–X (X = Cl, Br, I) created by accurate atomistic simulations that are more complete than those currently available in databases.