New Prompt Learning Boosts Crystal Property Prediction in GNNs

This research introduces an innovative soft prompt learning framework designed to enhance the accuracy of Graph Neural Networks (GNNs) in predicting crystal properties. Traditional GNNs often struggle with incorporating all relevant chemical and structural features, leading to larger models and reliance on deep domain expertise. The new framework addresses this by capturing latent features not explicitly fed into the GNN. The proposed method employs a multilevel approach, utilizing both node-level prompts to understand local chemical semantics and graph-level prompts to encode global structural symmetry. This lightweight framework integrates seamlessly with existing GNN encoders. Experiments demonstrate a substantial performance improvement, ranging from 3% to 15%, on various benchmark datasets. A key advantage of this prompt learning approach is its ability to facilitate knowledge transfer across different crystal properties. This means that models can achieve better prediction performance even for properties where training data is scarce, making it a versatile tool for materials science.