Materials characterization datasets often suffer from inconsistencies, especially historical collections of microstructural images that lack proper documentation and metadata. This presentation introduces AI-driven approaches to transform such microstructural microscopy data into well-organized resources for materials research.

We demonstrate self-supervised deep learning methods being developed within the AID4GREENEST project to automatically sort and curate large scanning electron microscopy datasets of steels by organizing them into a microstructural space. By analyzing the structure of this space, imperfections and imbalances in the datasets are deduced, revealing gaps in representation and potential biases in the collected data.

The case study presented shows how these methods successfully transform a historical metallurgical dataset into a valuable resource, enabling more efficient knowledge extraction and supporting new insights previously obscured by data inconsistencies.