Leveraging transfer learning for accurate estimation of ionic migration barriers in solids.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Ionic migration barrier (Em) is critical for applications like batteries, fuel cells, and sensors.
• Accurate estimation of Em is challenging, with previous methods relying on imprecise descriptors.
• Developing predictive models for Em is essential for accelerating materials discovery.

Purpose of the Study:

• To develop an efficient and accurate method for predicting the ionic migration barrier (Em) in diverse materials.
• To leverage transfer learning and graph neural networks for improved Em prediction.
• To establish a benchmark for machine learning models in predicting materials properties.

Main Methods:

• Utilized a graph neural network architecture with transfer learning principles.
• Pre-trained a model (MPT) on seven bulk properties and fine-tuned it on a dataset of 619 Em values.
• Incorporated architectural modifications to account for migration pathways and improve inductive bias.

Main Results:

• The best-performing fine-tuned model (MODEL-3) achieved a R² score of 0.703 ± 0.109 and MAE of 0.261 ± 0.034 eV on the test set.
• Demonstrated superior accuracy compared to classical machine learning, graph models trained from scratch, and machine learned interatomic potentials.
• Achieved 80% accuracy in classifying materials as 'good' ionic conductors.

Conclusions:

• Transfer learning strategies and MPT architectural modifications are effective for predicting Em.
• The developed model offers a significant advancement in accurately predicting ionic migration barriers.
• This approach can be extended to predict other data-scarce material properties.