Bayesian Framework for Inverse Inference in Manufacturing Process Chains

Process-property relations are central to ICME. Engineers are often interested in using these relations to make decisions on process configurations to achieve desired properties. This is known as the inverse problem and is typically solved using forward models (physics-based or data-based) in an optimization loop, which can sometimes be expensive and error prone, especially when used on process chains with multiple unit steps. We propose a Bayesian networks-based approach for modeling process-property relations that can be used for inverse inference directly. The solutions thus found can serve as good starting points for a more detailed simulation-based search. We also discuss how unit process models can be composed to do inverse inference on the process chain as a whole. We demonstrate this in a wire-drawing process where a wire is drawn in multiple passes to achieve desired properties. We learn a Bayesian network for a unit pass and compose it multiple times to infer process parameters of all passes together.