Convolutional Graph Neural Networks for Predicting Enthalpy of Formation in Intermetallic Compounds Using Continuous Filter Convolutional Layers

Accurately predicting the enthalpy of formation for intermetallic compounds plays a crucial role in materials design and optimization. This article proposes a novel deep learning approach for predicting formation enthalpy. This research develops a graph neural network combined with continuous filter convolutional layers to simulate quantum interactions between atoms. This enables direct learning from atom types and coordinates without simplifying into grid representations. This model demonstrates superior performance on the public JARVIS-DFT dataset compared to traditional machine learning methods. The introduction of continuous filter convolutional layers enhances the ability of graph convolutional neural networks to effectively learn atomic spatial features. This provides a new way to construct graph data structures from crystallographic information for materials science. Additionally, this work highlights the potential value of using graph convolutional neural networks to predict enthalpy of formation for intermetallic compounds.