Navigating Materials Space with ML-Generated Electronic Fingerprints

In the field of materials science, finding materials with specific properties is a major challenge due to the vastness of the search space, which makes random exploration prohibitively expensive. A more practical approach is to search for new materials within the proximity of known compounds that possess the desired property. In such an approach, fingerprinting methods are often used to measure the similarity of materials and group them into clusters. Such methods rely exclusively on the material’s structure to generate the fingerprint, which often does not correspond to clustering by desired property. To address this issue, electronic structure fingerprints that use properties such as the density of states (DOS) and band structure were proposed as an alternative. However, the computational cost of electronic structure calculations for tens of thousands of materials remains too high for rapid exploration. In this work, we developed a Graph Neural Network (GNN) ProDosNet which is trained on orbital Projected Density of States (PDOS) data and capable of predicting the electronic structure of materials at extremely low computational cost. With this model, we were able to generate PDOS fingerprints for all compounds present in the Materials Projects Database and cluster them by the similarity of their orbital PDOS, and therefore electronic properties. We demonstrate that using PDOS fingerprints allows finding materials that have similar electronic properties but drastically different structures.