Structure-Aware GNN-Based Deep Transfer Learning Framework For Enhanced Predictive Analytics On Small Materials Data

Published

January 2, 2024

Author(s)

Vishu Gupta, Kamal Choudhary, Brian DeCost, Francesca Tavazza, Carelyn E. Campbell, Wei-keng Liao, Alok Choudhary, Ankit Agrawal

Abstract

Modern data mining methods have been demonstrated to be effective tools to comprehend and predict materials properties. An essential component in the process of materials discovery is to know which material(s) (represented by their composition and crystal structure) will possess the desirable materials proper5 ties. For many materials properties, performing experiments and DFT computations are time-consuming and costly. Hence, it is challenging to build an accurate predictive model for such properties using conventional data mining methods due to the small amount of available data. We present a new framework for the materials property prediction task using crystal structure that leverages state of-the-art gnn-based architecture along with deep transfer learning techniques to drastically improve the model's predictive ability using the available experimental and/or DFT-computed datasets. We evaluated the proposed framework on 119 dataset-property combinations to find that the transfer learning models outperform the models trained from scratch in 109 cases, i.e., ∼ 92%. We believe that the proposed framework can be widely useful in accelerating materials discovery process in materials science.

Citation

Nature Machine Intelligence

Pub Type

Journals

Download Paper

https://doi.org/10.1038/s41524-023-01185-3

Local Download

Keywords

Deep transfer learning, material properties, density functional theory, materials informatics, graph neural network

Citation

Gupta, V. , Choudhary, K. , DeCost, B. , Tavazza, F. , Campbell, C. , Liao, W. , Choudhary, A. and Agrawal, A. (2024), Structure-Aware GNN-Based Deep Transfer Learning Framework For Enhanced Predictive Analytics On Small Materials Data, Nature Machine Intelligence, [online], https://doi.org/10.1038/s41524-023-01185-3, https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=935660 (Accessed October 17, 2025)