Holistic Machine Learning Approaches for Robust Microstructure Analysis: Insights from Phase Analysis and Grain Boundary Segmentation

With materials becoming more advanced and their microstructures becoming increasingly fine and complex, consisting of a combination of different phases or constituents with different substructures, existing characterization methods reach their limits regarding time efficiency and objectivity. This is where machine learning (ML)-based approaches apply, allowing an automatization of characterization methods with the ability to treat larger amounts and higher-dimensional data, bearing previously unused potential and enabling new analyses where satisfactory approaches used to be lacking.

The goal of ML training is to obtain a model that attains good results on previously unseen data, i.e., that is a powerful generalizer. ML has, in fact, already demonstrated its potential for a variety of tasks in microstructural analysis. However, the majority of publications deal with well-curated data sets exhibiting little variations – and a general understanding of robustness and generalization, as well as data set size and occurring variances, is therefore still lacking.

This presentation provides a holistic and comprehensive view of the ML implementation of two standard metallographic tasks, grain size determination and phase analysis, using the example of steel microstructures. Besides an objective ground truth, a deep understanding of the occurring variances and the model robustness, i.e., its stability against fluctuations and errors, is crucial. Variances and user influences – stemming mainly from specimen preparation, contrasting, and microscopy – are discussed and their impact on model robustness and generalizability analyzed. Additionally, the capabilities of fine-tuned data augmentation pipelines are discussed, and an outlook is given on how experimental and microscopic metadata can be incorporated into future ML models for robustness.