Researchers at the Indian Institute of Science (IISc) in collaboration with University College London have made significant progress in advancing the field of materials science through the innovative application of machine learning. Their breakthrough involves developing methods to predict material properties using limited datasets, which has the potential to transform the discovery and design of materials with desired characteristics, such as semiconductors and energy storage components.
In modern materials engineering, the prediction of properties such as electronic band gaps, formation energies, and mechanical resilience plays a crucial role in designing advanced materials. Machine learning models have emerged as powerful tools in this domain, offering the ability to analyze and predict properties based on data. However, the major challenge lies in the limited availability of high-quality data required to train these models. Experimental testing of materials to obtain property data is not only time-intensive but also expensive, creating a bottleneck for innovation.
Recognizing this limitation, a team led by Sai Gautam Gopalakrishnan, Assistant Professor in the Department of Materials Engineering at IISc, focused on addressing this challenge through a machine learning approach called transfer learning. Their study has unveiled a more efficient way to predict material properties by leveraging this method.
Transfer learning involves pre-training a model on a large dataset and then fine-tuning it on a smaller, task-specific dataset. The fundamental idea is to allow the model to gain a broad understanding of general features in the initial training phase, which can then be specialized for a specific task. Gopalakrishnan offers a simple analogy: “The model first learns to classify general images, such as identifying cats versus non-cats, and then adapts to a more specific task, like diagnosing cancer by distinguishing tissue samples with tumors from those without.”
Machine learning models operate by processing input data—such as images or material structures—and producing outputs based on learned patterns. The process begins with raw data input, which passes through layers in the model that extract and refine features. For example, initial layers might identify basic elements like edges, while deeper layers recognize complex patterns, such as shapes or structures. These refined features eventually lead to high-level classifications or predictions.
To develop a model suitable for material property predictions, the team employed Graph Neural Networks (GNNs). GNNs are particularly adept at handling graph-structured data, such as the three-dimensional crystal structures of materials. In a GNN, atoms in a material are represented as nodes, and the bonds between them as edges. This approach enables the model to process and understand the intricate relationships within a material’s structure.
The success of such models depends heavily on their architecture—the arrangement and number of layers, as well as how these layers are connected. The IISc team meticulously optimized the architecture and determined the required training data size for accurate predictions. Additionally, they pre-trained their model by fine-tuning only specific layers while keeping others “frozen.” This selective adjustment of layers allowed the model to retain broad knowledge while specializing in predicting material properties.
Using this optimized and pre-trained model, the team provided data on specific material properties, such as dielectric constants and formation energies, as inputs. The model then successfully predicted other critical properties, including the piezoelectric coefficient—a measure of a material’s ability to generate electric charge under mechanical stress.
One of the most remarkable findings was the superior performance of their transfer learning-based model compared to models trained from scratch. To enhance its versatility, the team employed a framework known as Multi-property Pre-Training (MPT). This method allowed the model to simultaneously pre-train on seven distinct properties of bulk 3D materials. Surprisingly, the model demonstrated the ability to predict the band gap values of 2D materials, even though it had not been specifically trained on such data.
The implications of this research extend far beyond academic curiosity. The team is now applying their model to predict ionic mobility within battery electrodes, a critical factor in improving energy storage devices. Faster ion movement in electrodes can lead to more efficient batteries with higher capacities and longer lifespans. Additionally, the model is being explored for its potential to enhance semiconductor development. By predicting the likelihood of point defect formation in semiconductors, researchers can identify materials that are more stable and efficient, which aligns with India’s strategic focus on advancing semiconductor manufacturing.
Gopalakrishnan highlights the broader impact of their work: “This approach can pave the way for better semiconductors and energy storage technologies, contributing to India’s technological and industrial growth.”
The study exemplifies the transformative potential of combining cutting-edge machine learning techniques with materials science. By overcoming the constraints of limited data, the researchers have demonstrated how intelligent computational tools can accelerate innovation in material discovery and design. As these models continue to evolve and adapt to new challenges, they hold the promise of reshaping industries and addressing global needs for advanced materials in technology, energy, and sustainability.
Reference: Reshma Devi et al, Optimal pre-train/fine-tune strategies for accurate material property predictions, npj Computational Materials (2024). DOI: 10.1038/s41524-024-01486-1