Physics-Based Synthetic Data Model for Automated Segmentation in Catalysis Microscopy.

Area of Science:

• Catalysis research
• Materials science
• Microscopy imaging

Background:

• Quantitative analysis of dynamic processes in catalysis is hindered by large microscopy datasets.
• Machine learning segmentation models require high-quality annotated training data, which is labor-intensive to produce.

Purpose of the Study:

• To develop a physics-based synthetic data generation model for training machine learning segmentation algorithms.
• To enable quantitative analysis of catalyst phase transitions using environmental scanning electron microscopy (ESEM) data.
• To gain insights into the phase transition of cobalt oxide catalysts during isopropanol oxidation.

Main Methods:

• A physics-based sequential synthetic data model was developed, substituting expert-annotated data.
• The model generates synthetic ESEM images by composing room-temperature catalyst images with dynamically evolving synthetic cracks.
• Crack growth mimics physical principles, avoiding pores and following surface paths observed in ESEM data.

Main Results:

• The synthetic data generation approach successfully approximates the catalyst phase transition.
• The physics-based method for crack formation demonstrates a reduced rate of false positives compared to random approaches.
• The trained neural network can perform semantic segmentation of ESEM data, facilitating analysis of the phase transition.

Conclusions:

• Physics-based synthetic data generation is a viable alternative to expert annotation for training machine learning models in catalysis.
• This approach facilitates the quantitative analysis of complex dynamic processes, such as catalyst phase transitions.
• The developed method aids in understanding catalyst deactivation mechanisms and optimizing catalyst performance.