Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 60: Topical Session: In Situ and Multimodal Microscopy in Materials Physics III

MM 60.2: Talk

Thursday, March 21, 2024, 16:00–16:15, C 130

Cracking Catalysts: A Synthetic Data Approach to Microscopy Image Segmentation — •Maurits Vuijk, Gianmarco Ducci, Luis Sandoval, Karsten Reuter, Thomas Lunkenbein, and Christoph Scheurer — Fritz-Haber-Institut der MPG, Berlin

In catalysis research, the amount of microscopy data acquired when imaging dynamic processes is typically too vast for non-automated segmentation. The challenge in developing a conventional automated process is that this requires more high-quality annotated training data than available in most cases. In our approach, we thus substitute expert-annotated data with a physics-based synthetic data model.

Our electron microscopy (environmental SEM) data is collected from the process of propanol oxidation to acetone over cobalt oxide. At a certain temperature during the reaction, a phase transition occurs and cracks form on the porous surface, reducing the selectivity of the catalyst. To generate synthetic image data that approximates this transition, our algorithm composes images of the pristine room-temperature catalyst with dynamically evolving synthetic cracks satisfying two physical construction principles. First, crack growth propagates along surface paths which avoid close vicinity to nearby pores. Second, each growing path successively widens and is rendered with increasing contrast to mimic depth over several frames. We then train a neural network model with the sequential data set to obtain a segmented time series of the collected data. This novel method can be used in real-time operation to guide the microscope in capturing the initial nuclei of phase transitions within the system.

Keywords: Electron microscopy; Machine learning; Computer vision; Synthetic Data; Catalysis