A foundational step in applying artificial intelligence to materials discovery is the ability to accurately classify materials based on desired quantum properties. We have taken a significant leap toward this goal by developing an advanced deep learning model to identify new topological materials with the potential to function at room temperature.

Our approach combines persistent homology with a graph neural network (GNN) to form a robust and interpretable classification framework. This model achieves a classification accuracy of 91.4% and an F1 score of 88.5%, outperforming existing state-of-the-art classifiers for distinguishing topological from non-topological materials.

What sets our model apart is its ability to generalize beyond the training data. It maintains high confidence when applied to out-of-distribution samples and newly discovered topological materials, a key advantage in the search for uncharted quantum phases.

The graph neural network component encodes the intricate atomic relationships within a crystal structure, efficiently handling the non-Euclidean geometry of materials using a relatively shallow network. Meanwhile, the integration of persistent homology introduces a topological analysis pipeline that captures shape and connectivity features of the crystal structure—boosting both model robustness and performance.

This classifier is now positioned as a powerful tool for high-throughput screening of topological materials. It significantly streamlines the discovery pipeline, paving the way for novel quantum materials

Link to the paper