Graph Prompt Learning Boosts Crystal Property Prediction in GNNs.

Graph Neural Networks (GNNs) have become a powerful tool for rapidly and accurately predicting various properties of crystals. However, these models often embed extensive domain-specific knowledge into their graph encoding modules, leading to larger parameter sizes and a heavy reliance on expert domain knowledge. Furthermore, explicitly incorporating all relevant chemical and structural features that influence a specific crystal property into a GNN encoder can be a challenging and incomplete task. To address these limitations, this research proposes a novel soft prompt learning framework designed to capture latent features crucial for property prediction that might not be explicitly provided to the GNN. The framework is multilevel, comprising both node-level and graph-level soft prompts. Node-level prompts capture local chemical semantics of different atom types, while graph-level prompts encode the global structural symmetry of the crystal graph. This prompt learning framework is lightweight and can be seamlessly integrated with any existing GNN encoder, making it highly versatile. Extensive experiments on popular benchmark datasets demonstrate that incorporating this prompt learning significantly improves the performance of state-of-the-art GNN models by 3% to 15% in crystal property prediction tasks. Additionally, the learned soft prompts facilitate cross-property knowledge transfer, which is particularly beneficial for enhancing prediction performance when training data for certain properties is limited.