Efficiently exploiting process-structure-property relationships in material design by multi-information source fusion

Abstract Materials design calls for the (inverse) exploitation of Process-Structure-Property (PSP) relationships to produce materials with targeted properties. Unfortunately, most materials design frameworks are not optimal, given resource constraints. Bayesian Optimization (BO)-based frameworks are increasingly used in materials design as they balance the exploration and exploitation of design spaces. Most BO-based frameworks assume that the design space can be queried by a single information source (e.g. experiment or simulation). Recently, we demonstrated microstructure-sensitive design of alloys with a BO framework capable of exploiting multiple information sources. While promising, the previous framework is limited as it assumes that the optimal microstructure is always feasible and considers microstructural features as the design space. Herein, we sidestep this unwarranted assumption and instead consider that chemistry and processing conditions constitute the design space amenable to optimization. We demonstrate the efficacy of our expanded framework by optimizing the mechanical performance of a ferritic/martensitic dual-phase material by adjusting composition/processing parameters. The framework uses thermodynamic results to predict microstructural attributes which are then used to predict the mechanical properties using a variety of micromechanical models and a microstructure-based finite element model. The final stage involves implementing model reification and information fusion, and a knowledge-gradient acquisition function to determine the next best design point and information sources to query. A detailed discussion of the various components and demonstration of how the framework can be implemented under three sets of cost-based constraints is presented.