Efficient machine learning of solute segregation energy based on physics-informed features

Machine learning models solute segregation energy based on appropriate features of segregation sites. Lumping many features together can give a decent accuracy but may suffer the curse of dimensionality. Here, we modeled the segregation energy with efficient machine learning using physics-informed features identified based on solid physical understanding. The features outperform the many features used in the literature work and the spectral neighbor analysis potential features by giving the best balance between accuracy and feature dimension, with the extent depending on machine learning algorithms and alloy systems. The excellence is attributed to the strong relevance to segregation energies and the mutual independence ensured by physics. In addition, the physics-informed features contain much less redundant information originating from the energy-only-concerned calculations in equilibrium states. This work showcases the merit of integrating physics in machine learning from the perspective of feature identification other than that of physics-informed machine learning algorithms.

Automated summary

In this study, the segregation energy was modeled using efficient machine learning with physics-informed features. These features outperformed the many features used in the literature and achieved a balance between accuracy and feature dimension. The strong relevance to segregation energies and the mutual independence ensured by physics contribute to their excellence.