ReLMM: Reinforcement Learning Optimizes Feature Selection in Modeling Materials.

Area of Science:

• Materials Science
• Machine Learning
• Computational Chemistry

Background:

• Identifying crucial physical features for material properties is vital for efficient materials discovery.
• Redundant features complicate the search domain, hindering optimal material design.
• Existing methods may not efficiently identify minimal, non-redundant feature sets.

Purpose of the Study:

• To introduce a reinforcement learning-based material model (ReLMM) for identifying minimal feature subsets.
• To improve the accuracy and efficiency of material property prediction by selecting optimal features.
• To analyze feature importance across different scales (molecular, mesoscale, device-scale).

Main Methods:

• Developed a reinforcement learning-based material model (ReLMM).
• Applied ReLMM to synthetic multiscale datasets for feature importance analysis.
• Compared ReLMM's performance against state-of-the-art feature selection tools (LASSO, XGBoost).

Main Results:

• ReLMM successfully identifies the relative importance of physical features across different scales.
• The model achieves prediction accuracy comparable to or better than existing methods.
• ReLMM demonstrates effectiveness in selecting near-minimal feature sets for predicting material properties like band gap.

Conclusions:

• ReLMM offers a powerful tool for efficient materials discovery by identifying optimal, non-redundant feature subsets.
• The approach aids in uncovering hierarchical material structures from multiscale data.
• ReLMM enhances predictive modeling for materials with specific target properties.