Chennai Mathematical Institute

Seminars




Seminar Announcement
Date: Wednesday, 20 August 2025
Time: 3.30 PM
Venue: Seminar Hall
Topology-Aware Graph Neural Networks: Enhancing Representation Learning for Materials Property Prediction

Sameer Deshpande
IIT Madras.
20-08-25


Abstract

Accurate prediction of material properties is essential for accelerating materials discovery and design. Graph neural networks (GNNs), particularly the MatErials Graph Network (MEGNet), have demonstrated strong performance in predicting the formation energies of crystals. But traditional message-passing GNNs often fail to capture critical topological structures within materials. In this work, we enhance MEGNet by integrating a topological graph layer (TOGL), which leverages topological data analysis (TDA) to compute persistent homology-based features for each crystal graph, enabling the model to extract and utilize global structural information. We train our topology-enhanced MEGNet model on datasets of diverse crystal representations. The incorporation of topological features significantly improves predictive accuracy. The TOGL layer, equipped with a learnable filtration function, captures essential topological features such as connected components and cycles, addressing limitations in standard GNN architectures. Our findings highlight the power of AI-driven topological feature extraction in advancing materials informatics. By integrating TDA with deep learning, we enhance the model’s ability to understand complex material structures, paving the way for more accurate and interpretable predictions in computational materials science. Our model achieves substantial performance improvements over existing benchmarks on the Materials Project and JARVIS datasets