Physics inspired Machine Learning of Eliashberg Spectral Function for High-Throughput Screening of Novel Superconductors

The Eliashberg spectral function (α²F) is at the core of understanding the electron-phonon superconducting properties of a material. But calculating α²F from ab initio methods for a large number of materials is expensive, impeding the high-throughput search of novel superconductors. With machine learning models gaining prominence over the past decade for material properties, our work harnesses this potential for α²F prediction. In this presentation, I will present our expansive database of theoretically calculated high-quality α²F and introduce our state-of-the-art equivariant neural network models trained using it. This talk also elucidates the nuances of training models for predicting continuous properties such as α²F, focusing on innovative physics-inspired node embeddings to bolster model performance. For the α²F-derived properties like the electron-phonon coupling λ, and two moments of the frequency (〈ω_log〉, and 〈ω²〉), our base model, relying solely on crystal structure, achieves a Pearson correlation coefficient of 0.3, 0.7, and 0.8. In comparison, integrating moderately expensive node embeddings increases the Pearson correlation coefficients of 0.5, 0.8, and 0.9 on the derived properties. An intriguing power-law relationship between model loss and training dataset size is observed, emphasizing the imperative for community-collaborated expansion of α²F databases.