Preview

PDF, English - main document Download (1MB) | Lizenz: Bayesian optimization framework for data-driven materials design by Tosato, Giovanna ; Koeppe, Arnd ; Selzer, Michael ; Nestler, Britta underlies the terms of Creative Commons Attribution 4.0

Abstract

The improvement of experimental design and the optimization of materials’properties with complex and partially unknown behaviors are common problems in material science. In the context of aqueous foams, the microstructure has a major influence on the properties of the resulting foam. Multiple interlinked parameters yield a large design space that requires tuning to tailor the microstructure evolution and resulting physical qualities. Our goal is a data-driven framework that uses machine learning to guide both experiments and simulations in an autonomous closed-loop. This iterative approach presents a valuable opportunity to accelerate materials development processes. A design of experiments methodology utilizing Bayesian Optimization is used to efficiently explore and exploit the search space, while minimizing the number of required evaluations. This approach allows to select the next most informative evaluation to perform, autonomously and adaptively learning from the already acquired data. The designed workflow is implemented into the data platform Kadi4Mat1, which provides the possibility of storing heterogeneous provenance data, along with a common workspace to integrate analysis methods and visualization. Our contribution within Kadi4Mat strongly relies on the reuse of data, and it is an example of the close interoperability between experimental and simulation research that the platform supports, in full alignment with the FAIR principles. Acknowledgements: This work is funded by the Ministry of Science, Research and Art Baden-Württemberg (MWK-BW) in the project MoMaF–Science Data Center, with funds from the state digitization strategy digital@bw (project number 57).