Speaker
Description
Material databases contain vast amounts of information, often harboring intricate connections and dependencies within material systems, some of which may remain undiscovered. Their structured organization naturally lends itself to the application of machine learning techniques. Through machine learning, we can unlock the tools necessary to discover potentially hidden structure-property relationships.
In this study, we present a use case where generic interactive workflows employ machine learning to automatically uncover such linkages from the research data infrastructure Kadi4Mat (https://kadi.iam.kit.edu/).
Within this infrastructure, data from simulated and experimental analyses of material systems are stored using a unified metadata scheme for coherent structuring.
In our use case, we first extract this uniformly structured data from the Kadi4Mat platform and prepare it for use in machine learning methods.
We then iteratively train neural networks to identify correlations between the microstructural compositions of the investigated materials and their resulting macroscopic properties. Using explainable AI techniques, in particular layer-wise relevance propagation, we identify the most influential parameters governing macroscopic properties. This allows us to streamline our neural network to focus only on key microstructural features to predict macroscopic properties.
Ultimately, we refine our network to predict macroscopic properties from minimal inputs, yielding a comprehensive material property map. This map concisely summarises the results of our network, allowing the macroscopic properties of a material to be quickly and easily determined from its microstructural composition.
This streamlined approach speeds up the materials research process and facilitates a data-driven accelerated development of new materials by providing researchers with invaluable insights into structure-property relationships.