Multi-task learning of solute segregation energy across multiple alloy systems

Machine learning can effectively model solute segregation energies at various sites within grain boundary networks. However, single-task learning approaches were usually used in the literature work, requiring as many learning tasks as there are alloy systems and reducing efficiency especially when dealing with numerous systems. In this work, we proposed a multi-task learning strategy to simultaneously learn segregation energies across multiple alloy systems within a hyperparameter-sharing architecture by integrating the data sets of these alloys. With appropriate algorithms and data integration methods, the best multi-task models can learn multiple alloy systems as accurately as the best single-task models. This simple strategy works due to the preserved correlation between site features and segregation energy after data integration, despite the change in how the correlation exists that renders some algorithms that work in single-task learning no longer effective in multi-task learning. This work provides an efficient and convenient approach to modeling site-specific solute segregation energy and will assist in the prediction of segregation behavior in alloy design.