Enhancing materials property prediction by leveraging computational and experimental data using deep transfer...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predictive modeling accelerates materials discovery by reducing the search space for Density Functional Theory (DFT) computations.
• Existing models inherit prediction errors from DFT computations and discrepancies with experimental data.

Purpose of the Study:

• To develop robust prediction models for material properties by leveraging deep transfer learning.
• To integrate large DFT-computed datasets, smaller DFT datasets, and experimental observations.

Main Methods:

• Applied deep transfer learning to combine data from the Open Quantum Materials Database (OQMD) and other sources.
• Developed a model to predict the formation energy of materials from their compositions.

Main Results:

• Achieved a mean absolute error (MAE) of [Formula: see text] eV/atom for formation energy prediction using an experimental dataset of [Formula: see text] observations.
• The proposed approach significantly outperforms existing machine learning (ML) prediction models based on DFT computations.
• The model's accuracy is comparable to the MAE of DFT computation itself.

Conclusions:

• Deep transfer learning offers a powerful strategy to build accurate predictive models for materials discovery.
• Integrating diverse datasets, including experimental observations, is crucial for overcoming limitations of purely computational predictions.
• This method enhances the reliability and efficiency of identifying promising new materials.