Predicting mechanical properties of newly generated two-dimensional materials with minimum uncertainty

Abstract

Two-dimensional (2D) materials exhibit exceptional properties. Thus, many studies have been conducted to discover novel 2D materials with unique characteristics or to find new ways of utilizing existing 2D materials. However, the existing open databases of 2D materials are often inefficient for this purpose. In this study, a material discovery framework is developed to identify new 2D materials using a deep learning-based generative model. First, a previous 2D database is adopted as a training set to develop a machine learning-based surrogate model for predicting the mechanical properties. Next, 2D candidates are generated, and their structural validity is confirmed by employing a classification model and checking their similarities to existing 2D materials. The uncertainty in the predicted mechanical properties of the generated materials is measured and the actual values are verified using density functional theory calculations. A total of 360 structures are newly identified according to the exploration method and the mean absolute error is significantly reduced from 206.025 to 10.185 N/m. We believe that the developed framework is general and can be further modified to search for novel 2D materials satisfying target physicochemical properties.