Advancing AI-Driven Microstructure Analysis through Correlative Microscopy Approaches

Artificial intelligence (AI)-based microstructure analysis has gained significant attention in recent years and is now a ubiquitous topic at scientific conferences. While AI has already proven its ability to automate various standard tasks, its true potential lies in enabling the analysis of highly complex microstructures for which no satisfactory evaluation methods previously existed.

One of the major challenges in studying such microstructures is the generation of a reliable ground truth. This process is not only time-consuming and labor-intensive but can also lead to inconsistencies due to subjective expert interpretation. To address these challenges, correlative microscopy approaches play a crucial role. Correlative microscopy refers to the combination of multiple complementary microscopy techniques, typically across different length scales, to analyze the same region of interest.

This talk is positioned at the intersection of traditional metallography and the emerging field of digital materials analysis. Various case studies will be presented to illustrate how correlative microscopy has been utilized to create robust datasets, enabling AI-driven microstructure analysis. These include fundamental investigations using light microscopy (LM) and scanning electron microscopy (SEM) to determine resolution limits and measurement uncertainties in LM-based evaluations, as well as combined LM, SEM, and electron backscatter diffraction (EBSD) analyses to distinguish different steel microstructures.

Furthermore, the presentation will discuss how correlative microscopy data can be leveraged to generate automated annotations for AI training, significantly reducing the need for manual labeling. As a final example, the use of generative AI to transform standard LM images into high-resolution SEM images will be demonstrated. In general, the increased effort associated with correlative microscopy is primarily intended for the creation of machine learning training datasets and reference data, with the ultimate goal of simplifying routine evaluations to the most basic microscopy method necessary.