Kadi4Mat: AI Driven Extraction of Structure-Property Linkages from Material Databases

When data is available in a structured format, machine learning can effectively identify complex
relationships and hidden dependencies. Materials databases are prime examples, providing rich and
well-organised datasets on different material systems. Applying machine learning to these databases
can therefore be a powerful tool for discovering hidden structure-property linkages.
In this study, we present a use case where generic interactive workflows use machine learning to
automatically reveal such connections within the research data infrastructure Kadi4Mat
[https://kadi.iam.kit.edu/].
Within this infrastructure, data from both simulated and experimental analyses of materials systems
are organised and structured using a standardised metadata scheme.
In our use case, we begin by extracting the uniformly structured data from the Kadi4Mat platform
and preparing it for machine learning applications.
We then iteratively train neural networks to detect correlations between the microstructural
compositions of the studied materials and their resulting macroscopic properties. By employing
explainable AI techniques, in particular layer-wise relevance propagation, we identify the most
... mehr
influential parameters that affect macroscopic properties. This enables us to refine our neural
network to focus only on the critical microstructural features for predicting macroscopic properties.
Ultimately, we optimise our network to predict macroscopic properties from minimal inputs,
generating a comprehensive material property map. This map is a concise summary of the results
from our network, and enables the quick and easy determination of a material's macroscopic
properties from its microstructural composition.
This streamlined approach accelerates the materials research process and supports data-driven
development of new materials by providing researchers with valuable insights into structure-
property linkages.