Pretraining Strategies for Structure Agnostic Material Property Prediction

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Machine learning (ML), particularly graph neural networks (GNNs), excels at predicting material properties.
• Traditional ML models often require computationally expensive, relaxed crystal structures.
• Structure-agnostic methods use fixed, hand-engineered descriptors, limiting learnability.

Purpose of the Study:

• To develop and evaluate novel pretraining strategies for structure-agnostic, learnable material property prediction frameworks.
• To enhance the performance of ML models that predict material properties without relying on relaxed crystal structures.
• To improve data efficiency and accuracy, especially for small datasets, in materials property prediction.

Main Methods:

• Proposed three pretraining strategies: self-supervised learning (SSL), fingerprint learning (FL), and multimodal learning (ML).
• Applied these strategies to the Roost architecture, a learnable, structure-agnostic framework.
• Evaluated the efficacy of pretraining on downstream material property prediction tasks.

Main Results:

• Demonstrated significant performance improvements on small datasets.
• Showcased enhanced data efficiency on larger datasets.
• Validated the effectiveness of the proposed pretraining strategies in leveraging unlabeled data.

Conclusions:

• Pretraining strategies significantly boost the performance of structure-agnostic ML models for material property prediction.
• These methods offer a powerful way to utilize unlabeled data, reducing the need for extensive labeled datasets.
• The approach holds great potential for accelerating materials discovery and design.