Robustness of Machine Learning Models in Microstructural Analysis

• Applicant: Professor Dr.-Ing. Frank Mücklich
• Subject Area: Thermodynamics and Kinetics as well as Properties of Phases and Microstructure of Materials
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 541968020

The examination and analysis of a material’s microstructure is at the same time the basis for the quality assurance of existing materials, their further development and the optimization of new materials with regard to their microstructure. The microstructure is regarded as a central information source that stores cross-scale information about various processing steps and also determines the mechanical and functional properties. Due to the increasingly complex and overall multiform microstructural constituents, conventional evaluation approaches, e.g., for microstructure segmentation and classification, are reaching their limits. Machine learning (ML)-based approaches must represent a fundamental and decisive tool for microstructure analysis in the future. They promise improved, more objective, and reproducible microstructure segmentations and classifications or even enable new evaluations. Currently, however, a fundamental understanding of generalization and robustness of ML models is still lacking but is necessary for ML to be used reliably and successfully in the long term and thus to be established as a fundamental tool for microstructural analysis and microstructure-based material development. Robustness and generalizability refer to variances as well as operator and equipment influences that occur during the individual steps leading to acquiring a micrograph (e.g., sample preparation, sample contrasting, image acquisition) and can have a significant influence on the appearance of the microstructure in the final image. The aim of the project is to establish ML as a robust and reliable tool for microstructural analysis. For this purpose, adapted and generally applicable strategies on how to build such models shall be developed and a material science understanding of robustness and variances shall be achieved. As a task, the classification of steel microstructures and the segmentation of multiphase microstructures of high chromium cast iron will be addressed. The following questions are to be answered. (i) Will a single and generalizing ML model trained with a large variance of data covering a wide range of experimental constraints, or specific ML models, each covering a particular set of constraints, be able to achieve better results. (ii) Furthermore, a materials science understanding of robustness shall be achieved by correlating the performance of ML models with microstructure features. This will involve investigating which variances have what influence, what the most significant experimental parameters are, and which variances can be controlled and how and with what effort. (iii) Furthermore, it will be investigated in which use cases and in which form metadata (in this case the recording of the experimental conditions) in the ML model can improve ML classification or segmentation.

• DFG Programme: Research Grants