Computational Materials Discovery

The Challenge:

The discovery of functional materials drives technological advancement, yet identifying materials with specific properties within vast chemical composition spaces remains challenging. Complex structure-property relationships make this search inherently difficult. Traditional approaches, including trial-and-error experimentation and high-throughput ab initio methods, are time-consuming and resource-intensive, limiting exploration across design spaces.

Our Innovation:

We use Machine Learning (ML) to accelerate materials discovery through data-driven approaches that identify patterns within chemical and structural datasets. ML models make rapid, physics-informed predictions about material properties, streamlining the discovery process and reducing computational costs.

To address interpretability limitations of “black-box” models, we develop hybrid ML approaches that integrate predictive power with physical interpretability. These methods map feature vectors to physical descriptors, ensuring models provide both accurate predictions and meaningful scientific insights.

Recent Achievements:

We designed ML models for materials tailored to quantum information processing, focusing on predicting defects analogous to nitrogen-vacancy (NV) centers in diamond—defect systems with electronic levels suitable for physical qubits. Our models achieve F1 scores exceeding 0.98 for classification tasks and Matthews correlation coefficients above 0.90 on imbalanced datasets while preserving interpretability.

These materials address challenges in building qubit arrays with long coherence times at room temperature.