Screening Cu-Zeolites for Methane Activation Using Curriculum-Based Training.

Area of Science:

• Computational heterogeneous catalysis
• Materials science
• Machine learning applications

Background:

• Machine learning (ML) combined with atomistic simulations accelerates catalyst discovery.
• Current ML models for catalysis lack interpretability and transferability, limiting their application.
• Zeolite catalysts are crucial for industrial reactions like methane activation.

Purpose of the Study:

• To develop interpretable and transferable machine learning models for zeolite catalyst screening.
• To systematically train reactive machine learning potentials (rMLPs) using a curriculum-based training (CBT) philosophy.
• To identify novel Cu-zeolite catalysts for methane activation.

Main Methods:

• Implemented a curriculum-based training (CBT) philosophy to train ML models.
• Combined diverse calculations to teach ML models about reactive potential energy surfaces.
• Screened thousands of [CuOCu]²⁺ sites across hundreds of Cu-zeolites using the developed rMLPs.

Main Results:

• Developed accurate, transferable, and interpretable reactive machine learning potentials (rMLPs).
• Identified MEI, ATN, EWO, and CAS as promising, previously unexplored zeolite structures for methane activation.
• Achieved high ensemble-averaged rates for [CuOCu]²⁺-catalyzed methane activation in identified zeolites.

Conclusions:

• The CBT philosophy effectively enhances ML model performance for catalyst discovery.
• This approach can be generalized to other zeolite-catalyzed reactions and heterogeneous catalysts.
• Significant progress in overcoming barriers in computational heterogeneous catalysis.