Improving crystal material property prediction with multi-view geometric graph transformer.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate representation of crystal structures is crucial for machine learning in materials simulations.
• Existing methods struggle to capture intricate geometric and topological features of crystals for property prediction.

Purpose of the Study:

• To develop a novel multi-view graph transformer (MGT) for enhanced crystal structure representation.
• To improve the accuracy and generalizability of crystal property prediction models.

Main Methods:

• Proposed MGT jointly models SE(3) invariant scalar and SO(3) equivariant directional representations.
• Utilized a mixture of experts inspired router for adaptive integration of embeddings.
• Employed multi-task self-supervised pretraining for model optimization.

Main Results:

• MGT achieved up to a 14% reduction in mean absolute error on crystal property benchmarks.
• Demonstrated significant performance improvements (up to 58%) in transfer learning tasks like catalyst adsorption energy and perovskite bandgap prediction.
• Ablation studies confirmed the contributions of pretraining and the router mechanism.

Conclusions:

• MGT is an effective and generalizable framework for crystal material property prediction.
• The approach shows significant potential to accelerate the discovery of novel materials.
• Jointly modeling invariant and equivariant representations is key to capturing crystal structure intricacies.