Entropy-based active learning of graph neural network surrogate models for materials properties

Area of Science:

• Computational materials science
• Machine learning applications in chemistry
• Data-driven materials discovery

Background:

• Graph neural networks (GNNs) are powerful tools for predicting material properties.
• Training GNNs typically requires large, labeled datasets, which are often scarce or expensive in materials science.
• This data limitation hinders the application of GNNs in many research areas.

Purpose of the Study:

• To develop an active learning scheme that reduces the amount of labeled data required for training GNNs.
• To enable GNNs to provide confidence measures for their predictions.
• To accelerate the discovery of new materials by efficiently selecting informative experiments.

Main Methods:

• Coupling a GNN with a Gaussian process to featurize solid-state materials.
• Predicting material properties along with a confidence score.
• Implementing an active learning strategy to guide data acquisition based on model uncertainty.

Main Results:

• The proposed scheme successfully predicts material properties with associated confidence levels.
• Active learning significantly reduces the amount of labeled data needed for model training.
• The active learning approach doubles the rate of performance improvement compared to random sampling.

Conclusions:

• Uncertainty quantification in GNNs enables efficient active learning for materials science.
• This approach can overcome data scarcity challenges, expanding GNN applicability.
• Accelerated training and data efficiency pave the way for new discoveries in data-limited materials research.