Speaker
Description
In materials development, creating new data points is often very costly due to the effort needed for materials synthesis, sample preparation and characterization. Therefore, all available knowledge in terms of data, physical models and expert knowledge should be exploited in the most efficient way (optimal knowledge exploitation). Moreover, the number of new samples/data points to be produced in terms of synthesis and characterization of new materials should be kept at a minimum to save time and money (sample efficiency). The ALPmat is an Active Learning Platform for MATerials design, targeted at optimal knowledge exploitation and sample efficiency. For optimal knowledge exploitation, a hybrid approach is followed, where physical models and expert knowledge are combined with data from observations. The resulting hybrid models are used for an Active Learning Loop (ALL) to improve the addressed properties in an iterative way via optimization of the material’s chemistry and processing conditions, while also minimizing the number of new samples ensuring sample efficiency. We will present the details of the ALPmat in terms of hard- and software for the platform backbone, the FAIR database, the framework for running physical modeling and Bayesian optimization algorithms, and integrated software services. Moreover, we will present data models for our use cases. These data models are linked to the state of the investigated sample and are used to uniquely identify use case-specific synthesis, processing and characterization steps. They serve as a basis for the ingestion of comprehensive metadata and ensure that the data are fully FAIR. Finally, we will show the first results for the application of the developed methodology and infrastructure for the use case of bainitic steels. For this use case, we are performing a multi-objective optimization of the uniform elongation and the yield strength as a function of chemical composition and processing conditions.