Fused Convolution and Attention Graph Neural Networks (FCAtGNN) Accelerates Eliashberg Spectral Function Predictions for Superconductors

Density functional theory (DFT) calculations combined with deep learning algorithms is a growing field of research for accelerating the discovery of materials with desirable properties. Recently, DFT calculations have been done on conventional superconductors through the Migdal-Eliashberg theory that explains superconductivity by electron-phonon coupling in the materials. The Eliashberg spectral function (electron-phonon spectral function) is a key part in the Migdal-Eliashberg theory to calculate the critical temperature (T_c) in conventional superconductors. Having of over 7,000 materials, we scale such that all the values are between 0-1 having the scaling factor as the integration of . We train Fused Convolution and Attention Graph Neural Networks (FCAtGNN) to predict the scaled with its scaling factor. We find that the predicted that comes from the scaled with its scaling factor yields lower MAE than directly predicting . After predicting then descaling , we calculate electron-phonon coupling parameter (λ) and the logarithmic average frequency (ω_log) to calculate T_c.