A Domain-Driven Approach to Metamodeling in Additive Manufacturing

Recent studies have shown advantages to utilizing metamodeling techniques to mimic, analyze, and optimize system input-output relationships in Additive Manufacturing (AM). This paper addresses a key challenge in applying such metamodeling methods, namely the selection of the most appropriate metamodel. This challenge is addressed with domain-specific AM information, derived from physics, heuristics and prior knowledge of the process. Domain-specific input/output models and their interrelationships are studied as a basis for a domain-driven metamodeling approach in AM. A metamodel selection process is introduced that evaluates global and local modeling performances, with different AM datasets, for three types of surrogate metamodels (polynomial regression (PR), Kriging, and artificial neural network (ANN)). A salient feature of this approach is its ability to seamlessly integrate domain-specific information in the model selection process. The approach is demonstrated with the aid of a metal powder bed fusion (PBF) case study and the results are discussed.