Semi-supervised Learning in Graph Neural Networks for Structural and Property Prediction Applied to Advanced Functional Materials Design

Machine learning is becoming an integrating part of computational materials science, being used to predict materials properties, accelerate simulations, design new structures, and predict synthesis routes of new materials. But its efficacy is undermined by problems of data scarcity and portability challenges. This work explores the potential of graph neural networks in developing a unified predictor for material properties. The goal is to create a versatile molecular model using atomic number and relative distances as exclusive features. The model aims to handle diverse molecular classes, scales, and theory levels, enhancing precision in predicting material properties, even with limited data. To achieve this, inspired by recent advances in Natural Language Processing, we propose a Masked Molecular Modeling task, training the model in a semi-supervised manner without explicit labels. This task allows the model to predict the atomic type of masked atoms in a molecular structure, giving the opportunity to aggregate diverse data sources and mitigating data scarcity issues. We also assess the capacity of the model to perform property prediction, even with masked elements, and compare it with state-of-the-art approaches. By incorporating a graph attention mechanism, we not only enhance the model’s performance but also gain valuable insights into its internal representation and processing. This contributes to meaningful explanations and a deeper understanding of the model’s workings.