Designing spinodoid architected materials by Bayesian optimization
Please login to view abstract download link
In this contribution, we propose a general framework to inversely designing structures using structure-property linkages. Typically, large datasets are necessary. Experiments alone are prohibitively expensive. Therefore, computational augmentation is employed to allow for data-driven approaches even in this data scarce regime. In an iterative procedure (1) mesostructures are characterized by descriptors, (2) effective properties are derived from numerical simulations, (3) structure-property linkages are set up using a Gaussian process, (4) descriptors of new candidate mesostructures are proposed by Bayesian optimization and (5) mesostructures are reconstructed. Steps 2 through 5 are repeated until a desired convergence criterion is reached, e.g., the uncertainty of the structure-property linkage is decreased or a mesostructured with preferable properties is found. This framework is applied and presented at the example of spinodoid structures [1]. Augmenting a small initial data set by in silico reconstructed spinodoid structures and their simulated effective properties allows for deriving improved structure-property linkages and, thus, finding potentially optimal structures or predicting properties. [1] Raßloff, Alexander, Paul Seibert, Karl A. Kalina, and Markus Kästner. 2024. “Inverse Design of Spinodoid Structures Using Bayesian Optimization.” doi:10.48550/arXiv.2402.13054.
