Heterogeneous Transfer Learning for Design and Manufacturing Process Modeling Through Input Mapping and Multi-Fidelity Gaussian Process

Abstract Data-driven modeling and optimization have become a cornerstone in several engineering design and modeling endeavors. Data-driven modeling frameworks enable faster optimization cycles and push the boundaries on the performance envelope of the system under study. Even with extended access to computational or manufacturing capabilities, the time\economic cost of evaluating design points are high when the simulation\experiment get closer to ground truth representation. The domain of transfer learning seeks to elevate this constraint on data requirements by leveraging data from other relevant or legacy data sets. However with the strong differentiation in design and manufacturing process, situations arise where the inputs of the target and source domain are heterogeneous in nature. Towards solving this problem, a heterogeneous transfer learning modeling framework is identified and implemented based on input mapping and multi-fidelity Gaussian processes. The framework is applied on three cases studies — (1) cantilever beam (2) ellipsoidal void and (3) friction stir welding. Each of the three case studies represent a different facet of the design engineering community at large (1) Design Differentiated (2) Complexity Differentiated (3) Manufacturing Modality. Key insights are provided on the input mapping transformation between the source\target domain and the performance of the integrated model. The results demonstrate adding source domain data facilitates a more accurate model.