Accelerating material property prediction using generically complete isometry invariants.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning accelerates periodic material and crystal property prediction, offering an efficient alternative to classical simulations.
• Representing unbounded periodic crystals is challenging, unlike finite molecules or proteins, requiring specialized approaches for machine learning algorithms.

Purpose of the Study:

• To adapt the Pointwise Distance Distribution (PDD) as a robust representation for periodic crystals in machine learning.
• To develop and evaluate a transformer model integrating PDD with compositional information for enhanced crystal property prediction.

Main Methods:

• Adapted the Pointwise Distance Distribution (PDD), a continuous and complete isometry invariant, for periodic crystal representation.
• Developed a transformer model incorporating a modified self-attention mechanism combining PDD with spatial encoding for compositional information.
• Validated the model on Materials Project and Jarvis-DFT databases.

Main Results:

• The PDD successfully distinguished over 660,000 periodic crystals in the Cambridge Structural Database based on their structure alone.
• The developed transformer model achieved accuracy comparable to state-of-the-art methods on crystal property prediction tasks.
• The PDD-based model demonstrated significantly faster training and prediction times compared to existing approaches.

Conclusions:

• The Pointwise Distance Distribution (PDD) provides an effective and computationally efficient representation for periodic crystals in machine learning.
• This approach enhances the speed and accuracy of crystal property prediction, facilitating faster material discovery and design.