Researchers at the Indian Institute of Science (IISc), with collaborators at University College London, have developed machine learning-based methods to predict material properties even with limited data. This can aid in the discovery of materials with desired properties, such as semiconductors.
In recent years, materials engineers have turned to machine learning models to predict which types of materials can possess specific properties such as electronic band gaps, formation energies, and mechanical properties, in order to design new materials. However, data on material properties – which is needed to train these models – is limited because testing materials is expensive and time consuming. This prompted researchers led by Sai Gautam Gopalakrishnan, Assistant Professor at the Department of Materials Engineering, IISc, to work on addressing this challenge. In a new study, they have found an efficient way to use a machine learning approach called transfer learning to predict the values of specific material properties.
In transfer learning, a large model is first pre-trained on a large dataset and then fine-tuned to adapt to a smaller target dataset. “In this method, the model first learns to do a simple task like classifying images into, say, cats and non-cats, and is then trained for a specific task, like classifying images of tissues into those containing tumours and those not containing tumours for cancer diagnosis,” explains Gopalakrishnan.
Machine learning models process input data, such as an image, and generate outputs, like identifying the shapes present in the image. The first layer of the model takes in the raw image input. Subsequent layers extract features from the image, such as edges, which are progressively refined. The final layers combine these features to recognise and classify higher-level features, like shapes. These models can be built using various architectures, such as Graph Neural Networks (GNNs), which work with graph-structured data like the three-dimensional crystal structure of any material. In GNNs, information in each layer is represented as nodes (atoms in a structure) and the connections between the nodes are represented as edges (bonds between atoms). For the current study, the research team developed a GNN-based model.
The architecture of the GNN, such as the number of layers and how they are connected, determines how well the model can learn and recognise complex features in the data. The team first determined the optimal architecture needed for the model and the training data size required for predicting material properties. They also pre-trained the model by tuning only some layers while “freezing” the others, explains Reshma Devi, first author and PhD student at the Department of Materials Engineering. To this optimised and pre-trained model, they provided data on material properties such as dielectric constant and formation energy of the material as the input, so the model could predict the values of specific material properties, like the piezoelectric coefficient.
The team found that their transfer learning-based model, which was first pre-trained and then fine-tuned, performed much better than models that were trained from scratch. They also used a framework called Multi-property Pre-Training (MPT) in which they simultaneously pre-trained their model on seven different bulk 3D material properties. Remarkably, this model was also able to predict the band gap value for 2D materials that it was not trained on.
The team is now using this model to predict how quickly ions can move within electrodes in a battery, which can potentially help build better energy storage devices. “It can also be used to make better semiconductors by predicting their tendency to form point defects, which can contribute to India’s push towards the manufacture of semiconductors,” Gopalakrishnan adds.
