Robust Lightweight Graph Neural Network Framework for Accelerating Crystal Structure Prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Graph Neural Networks (GNNs) are increasingly used for crystal structure prediction (CSP).
• Existing GNN-based CSP frameworks face limitations in robustness and computational efficiency.
• Sensitivity of GNN models to weight initialization is a critical but often overlooked issue in CSP.

Purpose of the Study:

• To develop a robust and computationally efficient GNN-based framework for crystal structure prediction.
• To address the sensitivity of GNNs to weight initialization and improve model selection.
• To enhance the performance of GNNs for CSP through data augmentation and pretraining strategies.

Main Methods:

• Employed a derivative-free optimization method for structural search.
• Utilized a supervised Graph Neural Network (GNN) as the energy evaluator.
• Introduced a model selection framework to identify appropriate GNN models for CSP.
• Implemented a data augmentation strategy using unrelaxed structures.
• Explored unsupervised GNN pretraining with and without augmentation.

Main Results:

• Developed a model selection framework to consistently identify suitable GNN models for CSP.
• Demonstrated that data augmentation with unrelaxed structures improves GNN performance.
• Showcased that unsupervised pretraining can enhance GNN-based CSP.
• Achieved performance comparable to complex GNNs using a lightweight CGCNN architecture.
• Validated the framework's effectiveness and computational efficiency.

Conclusions:

• The proposed framework offers a robust and computationally efficient approach to crystal structure prediction.
• The developed methods for GNN model selection and data augmentation are generalizable.
• This work paves the way for novel and high-throughput crystal structure prediction.
• Lightweight GNN architectures like CGCNN can achieve competitive performance in CSP.
• The findings contribute to advancing machine learning applications in materials science.