In recent years, artificial intelligence (AI) and supervised machine learning (ML) have also become ubiquitous in experimental materials science, demonstrating their enormous potential. In addition to being used for simple, tedious, and labor-intensive tasks, the focus is now also on the analysis of complex microstructures where traditional methods reach their limits. However, when really highly complex microstructure morphologies are studied and must be classified - the clean, objective, and reproducible ground truth assignment for supervised ML is crucial for implementing a sustainably successful ML-based microstructure analysis.

The essential basis is our fundamental knowledge of all steps in a completely reproducible preparation, contrasting, and imaging as well as a quantitative understanding of all parameters of microstructure analysis.

In this talk, exemplary examples will be presented where advanced microstructure characterization forms the indispensable basis for the successful application of ML for microstructure classification with very high accuracy.

1. Our novel "in-situ etch cell" allows the exploration of optimized and exactly reproducible microstructural contrasts of lowest variance in the micrographs. Strong ML models can thus be trained with much less effort of data.

2. correlative microscopy approaches allow objective assignment of complete ground truth by linking between different microscopy techniques including electron backscattering data.

3. Finally, definitions of classes for ML models are discussed. Classification based on various morphological features in series of 2D images need to be reviewed.

The success of such advanced use of ML for complex microstructure classification is impressive in terms of accuracy (i.e. recall as well as precision) compared to human expert’s observation and decision.