Molecular Property Prediction Toward Materials Science Applications Using Quantum-Inspired Graph Neural Networks

In the field of quantum chemistry and materials science, where new models can significantly speed up drug discovery, catalyst design, and material innovation, material property prediction is an important frontier. In this study, a Quantum-Inspired Graph Neural Network (QIGNN) model has been proposed to include geometric features and message passing for the prediction of quantum chemical interactions of molecules. The model utilizes the learned edge embeddings based on interatomic distance and direction vectors, followed by a multi-head regression for simultaneous prediction of multiple quantum properties. The model is evaluated on QM9 dataset which consists of 19 molecular properties. Using this model, mean absolute errors (MAEs) achieved are 0.1609 D for dipole moment, 0.4275 a03 for polarizability, 0.0803 eV for HOMO, 0.0960 eV for LUMO, and 0.1186 eV for the HOMO-LUMO gap. Energy-related properties such as U0, U, H, G achieve remarkably low MAEs of 0.0007 eV, while rotational constants (A, B, C) achieve errors below 0.1 GHz. Overall, the results show QIGNN’s ability to capture thermodynamic and energetic properties of systems, establishing it as an important model for catalysts, drug designing, material design, and chemical informatics