Automated Pynta-Based Curriculum for ML-Accelerated Calculation of Transition States.

Area of Science:

• Computational chemistry
• Catalysis
• Materials science

Background:

• Microkinetic models (MKMs) are crucial for understanding catalytic reactions but rely on computationally intensive DFT calculations.
• Identifying transition states for elementary reactions is a major bottleneck in MKM construction.

Purpose of the Study:

• To develop an automated workflow for training reactive machine learning potentials (rMLPs) to accelerate transition state calculations for MKMs.
• To improve the efficiency and accuracy of catalyst design using machine learning.

Main Methods:

• Utilized the Pynta kinetics workflow tool to automate iterative rMLP training.
• Applied the workflow to the silver-catalyzed partial oxidation of methanol, calculating 53 transition states.
• Investigated single vs. multiple rMLP models ('reaction class' approach) and fine-tuned pretrained foundational models.

Main Results:

• Achieved a 7x speedup in transition state calculations compared to DFT-only methods using a single rMLP.
• Demonstrated that a 'reaction class' approach with multiple rMLPs overcomes limitations of single models.
• Fine-tuning a pretrained graph neural network potential yielded a 20x speedup with 89% success rate.

Conclusions:

• The automated Pynta-based workflow significantly accelerates the construction of microkinetic models.
• Machine learning, particularly using multiple specialized rMLPs or fine-tuned foundational models, offers a powerful synergistic approach to advance catalysis research.