MXene Property Prediction via Graph Contrastive Learning

MXenes are an important class of 2-D materials with expected novel properties and myriad applications, including efficient energy conversion in batteries and solar cells, environmental and water treatment, supercapacitors, electromagnetic interference shielding, and catalysts. Unfortunately, MXenes are challenging to synthesize in the laboratory, and numerical models for simulating MXene properties, such as density functional theory (DFT), remain computationally expensive. For this reason, multiple prior works develop machine learning (ML) methods to rapidly predict MXene properties. However, these methods do not take into account the graphical structure of the MXene’s molecular bonds, reducing their accuracy. In this work, we introduce a novel ML method, based on Graph Neural Networks (GNNs), which includes MXene molecular graph information to predict electrical and physical properties. Our method transforms graph datasets to lower-dimensional embedding spaces with graph contrastive learning (GCL), a type of GNN, enabling high-accuracy predictions for downstream models trained with these embeddings. Our method obtains the following root mean square errors (RMSEs) when predicting properties from a DFT-based MXenes dataset (including many MXene properties not previously predicted with deep learning models): 0.418 J for Work Function, $0.518 \mathrm{eV}^{-1}$ for Density of States at Fermi Level, $\mathbf{0. 0 6 7}$ for Poisson’s Ratio, $21.65 \mathrm{~N} / \mathrm{m}$ for Bulk Modulus, 14.56 N/m for Shear Modulus, 35.66 N/m for Young’s Modulus, and 82% accuracy for Magnetic (T/F). Our method obtains lower mean absolute error (MAE) when predicting four target properties from the molecular QM9 dataset compared to a similar-complexity baseline. Predicting MXene properties more accurately, efficiently, and effectively promises to better enable MXene commercialization.