Latent Variable Machine Learning Framework for Catalysis: General Models, Transfer Learning, and Interpretability.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) is widely used for materials screening but often lacks generality and transferability.
• Existing ML models require retraining for each new application, especially in catalysis with numerous variables.
• This limits efficiency and broad applicability in discovering new materials.

Purpose of the Study:

• To develop a novel ML framework for general, interpretable, and reusable materials screening models.
• To enhance data efficiency and enable transfer learning across diverse applications.
• To incorporate fundamental chemical principles into the ML architecture.

Main Methods:

• Developed a new ML architecture utilizing latent variables to create specialized submodels.
• Integrated fundamental chemical principles, treating elements as discrete entities.
• Employed latent variables for physical interpretability and as feature representations for transfer learning.

Main Results:

• Achieved simultaneous prediction of adsorption energies on alloy surfaces with mean absolute errors (MAEs) of 0.20-0.25 eV.
• Demonstrated efficient transfer learning, creating accurate models with <10 data points.
• Showcased transfer learning to an experimental dataset with MAE < 0.15 eV.
• Validated robustness with heterogeneous and multifidelity datasets.

Conclusions:

• The proposed ML framework offers general, interpretable, and reusable models for accelerated materials discovery.
• The architecture enhances data efficiency and facilitates rapid model development through transfer learning.
• This approach significantly improves convenience and efficiency for researchers in computational materials science.