Artificial Intelligence (AI) and machine learning have become essential tools in the repertoire of many materials scientists and engineers, most recently fueled by the successes of large language models trained on enormous datasets. However, for day-to-day materials research, the data exchange between data-generating and data-analyzing collaborators is often handled through shared network/cloud directories, email attachments, or even the physical exchange of hard drives.
This process limits the scope of materials datasets, and a substantial amount of information about the data needs to be transferred in other ways.

The data-management platform Kadi4Mat allows storing, linking, and sharing data with the accompanying information saved as metadata. By linking dataset records and workflows that help describe and preprocess datasets for AI implementation, the interface KadiAI enables the development of "data-integrated AI" solutions for materials science and engineering. Data-integrated AI connects raw datasets provided by data-generating materials scientists and data-analyzing AI experts through serializable data, model, and training definitions.

In this overview, we demonstrate, how the Kadi ecosystem realizes data-integrated AI in multiple use cases. First, data-driven properties enrich the characterization of materials and enable the generation of variant microstructures. Second, intelligent interfaces in numerical simulation models learn to solve nonlinear spatiotemporal many-to-many mappings with alternating equilibrium states. Finally, explainable AI applied to materials datasets yields physics-explaining neural networks that complement physics-informing approaches to design specialized neural networks.