A Novel Graph Machine Learning Model for Structure-to-Property Prediction

Graph message-passing neural networks (MPNNs) are widely used to extract molecular and crystal structure representations, crucial for understanding structure-property relationships in materials discovery, molecule design, computational chemistry, and physics. Principal Neighbourhood Aggregation (PNA), an advanced MPNN, employs multiple aggregators and degree-based scalers for messages from an atom's neighborhood. However, using degree-based scalers may inadequately represent an atom's interactive environment, resulting in poor predictions for molecule and crystal structure-based properties. To address this, we propose adaptive-ChemGNN, an MPNN tailored for the atom's local interactive environment. It utilizes multiple adaptively weighted aggregations to effectively characterize atoms' local interactive environments. The adaptive aggregation weights are functions of neighboring atoms' aggregated messages and are learned during model training. Experimental results demonstrate that adaptive-ChemGNN accurately predicts diverse chemical properties and atomic features, such as the formation energy of high entropy metal oxides, molecular properties of the QM9 dataset, and force fields in molecular dynamics simulations. These findings underscore the model's significant potential in AI-driven material discovery, molecule design, MD simulations, and more beyond.