Machine learning method developed by IISc and UCL researchers uses transfer learning to predict material properties, paving the way for advanced semiconductors and energy storage solutions.
In a significant advancement for materials science, researchers at the Indian Institute of Science (IISc), in collaboration with University College London (UCL), have harnessed machine learning (ML) to predict material properties with minimal data. This innovation promises to accelerate the discovery of materials for applications like semiconductors and energy storage.
Materials engineers have increasingly turned to ML to predict properties such as electronic band gaps, formation energies, and mechanical strengths, enabling the design of novel materials. However, limited experimental data—a result of high testing costs and time constraints—has posed a challenge.
The team tackled this hurdle using a method called transfer learning. This approach pre-trains a model on a large dataset and fine-tunes it for a smaller, specific dataset. For example, a model trained to classify cats and non-cats can be adapted to classify tumor versus non-tumor tissue.
Their study employs Graph Neural Networks (GNNs), a sophisticated architecture suited for graph-structured data like 3D crystal structures. Here, atoms represent nodes, and bonds represent edges. By optimizing the GNN architecture and pre-training some layers while freezing others, the researchers built a model capable of predicting material properties like dielectric constants and piezoelectric coefficients.
Their Multi-property Pre-Training (MPT) framework allowed the model to learn from seven bulk material properties and successfully predict the band gap for 2D materials—a capability not explicitly trained for.
This holds promise for practical applications. The team is exploring the model’s use in predicting ion mobility in battery electrodes, crucial for advancing energy storage. Additionally, it can aid in manufacturing defect-resistant semiconductors, aligning with India’s semiconductor production goals.
