Task Agnostic Model for Material Property Prediction Using Joint Embedding Predictive Architecture

Task-agnostic models (TAM) are a promising approach to develop machine learning models that are both accurate and computationally efficient. Despite the complexities involved in their development, TAMs have the potential of revolutionizing computational chemistry by implementing generalized representations capable to predict multiple material properties. This paper, introduces a self-supervised graph-based representation for material structures, designed to be task-agnostic. We demonstrate that our representation, when combined with a linear model, is capable to predict a diverse set of properties. Our model’s performance is benchmarked using the publicly available Materials Project dataset. Additionally, our findings reveal that the model learns representations with sufficient expressive power, suitable for both regression and classification tasks. We demonstrate the versatility of our proposed model by predicting the total energy of cubic perovskites and the formation defect energy of impure crystal. Importantly, we demonstrate that our trained model is readily applicable to specialized, curated datasets , even those with a limited number of data points.