Data-efficient, simulation-based sample generation for the modelling of process-structure-property relations using machine learning
In manufacturing, the optimization of process-structure-property relations is driven by the demand for more and more specific and individually designed products and their performance requirements. Recent approaches in the optimization of process-structure-property relations make use of machine learning methods, which model the knowledge embedded in numerical simulations. The present contribution addresses the starting point of such approaches, namely the generation of data via numerical simulation. The simulations are governed by deterministic models that represent the underlying physical relations of interest. In metal forming, the relations relevant for work piece properties are those between a) forming processes or process-paths and microstructures, and b) between microstructures and material properties. Microstructure-property relations can be modeled by generalizing simulation results using machine learning approaches. The quality of the machine learning models to approximate these relations is highly depending on the generated data, hence on the sampled input space of the physical models. In the contribution presented, we analyze different sampling methods, such as space-filling design methods and active learning. The former are static sampling methods that a-priory define the locations where to evaluate the physical model, the latter describes sequential sampling strategies that are based on machine learning, in which the learner itself queries for data at optimal locations in terms of efficiency. The sampling methods are applied to single-phase metallic microstructures with an emphasis crystallographic texture. The aim is to generate sets of discrete orientations in order to sufficiently cover the space of macroscopic properties of interest.