Leveraging Transfer Learning and Chemical Principles toward Interpretable Materials Properties.

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Machine learning (ML) models offer powerful tools for understanding chemical properties and material performance.
• However, the complex, "black-box" nature of many ML models hinders interpretation and lacks physical foundations.
• This opacity limits trust and adoption in scientific applications.

Purpose of the Study:

• To develop a transfer learning strategy that incorporates fundamental chemical principles for interpretable ML models.
• To investigate the physical insights gained from interpreting ML models for inorganic compound formation energies.
• To address the limitations of current ML approaches in materials informatics.

Main Methods:

• Applied transfer learning by leveraging core chemistry principles to interpret ML models.
• Analyzed the performance of deep neural networks (DNNs) and recurrent neural networks (RNNs) with attention mechanisms for predicting formation energies.
• Focused on the ability of models to capture interelemental patterns and chemical relationships.

Main Results:

• The study identified deficiencies in DNNs for handling interelemental chemical patterns.
• Demonstrated that RNNs with attention mechanisms provide a more accurate abstraction of chemical relationships.
• Successfully predicted the periodic table structure with high precision using the interpretable ML approach.

Conclusions:

• The proposed transfer learning strategy enhances the transparency of ML models in materials science.
• Interpretable models are crucial for building trust and advancing the field of materials informatics.
• This work offers a pathway toward physically grounded and transparent ML solutions for chemical discovery.