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

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Machine learning models predict solute segregation energy using site features.
• Lumping features can lead to the curse of dimensionality, impacting model performance.
• Existing methods often lack a balance between accuracy and feature set size.

Purpose of the Study:

• To develop efficient machine learning models for predicting segregation energy.
• To identify and utilize physics-informed features for improved model performance.
• To demonstrate the advantage of integrating physical understanding into feature selection for machine learning.

Main Methods:

• Modeled segregation energy using machine learning with physics-informed features.
• Features were identified based on fundamental physical understanding of segregation sites.
• Compared performance against literature features and spectral neighbor analysis potential (SNAP) features.

Main Results:

• Physics-informed features achieved a better balance between accuracy and feature dimension.
• These features outperformed existing methods across different machine learning algorithms and alloy systems.
• The selected features showed strong relevance to segregation energies and mutual independence due to their physical basis.

Conclusions:

• Integrating physics into feature identification enhances machine learning model efficiency and accuracy for segregation energy prediction.
• Physics-informed features reduce redundant information compared to energy-only calculations.
• This approach highlights the value of physics-guided feature engineering in machine learning applications.