Explainable ensemble learning to predict anisotropic nanomaterial band gap using atomic-scale structural descriptors

Abstract

Predicting the electronic band gap of nanomaterials is essential for discovering and developing novel nanostructures with tailored properties for a myriad of applications, including biomedical and pharmaceutical applications. Band gap predictions are commonly performed using computational modeling approaches such as molecular dynamics simulations and density functional theory calculations. However, the high computational cost and extensive infrastructural requirements of these methods have impeded their wider adoption and consequently, more rapid and efficient discovery of high-performance nanomaterials. In this contribution, we demonstrate the use of explainable ensemble supervised learning to accelerate the prediction of the electronic band gap of anisotropic nanomaterials. We systematically assess the capacity of several base models and a stacking model in predicting the band gap of more than 300 polyhedral nanomaterials with varying atomic-scale structural attributes. By coupling ensemble learning with explainable feature selection, we achieve outstanding performance in predicting nanomaterial band gap, with R² values above 0.96 and MSE below 0.004. We anticipate that this work can further catalyze the development of machine learning and other artificial intelligence approaches to streamline the prediction of the band gap and other electronic properties of nanomaterials.