Probabilistic knowledge transfer of melt pool properties between wire- and powder-based laser-directed energy deposition

Laser wire– and laser powder–directed energy deposition, i.e., LW-DED and LP-DED, share similar working principles but suit distinct application scenarios. However, their process parameter spaces and the associated physical interactions are not identical. Hence, data from one process cannot be directly reused to predict the deposit characteristics of the other process when new feedstocks are employed. Therefore, this research aims to propose data-driven models that transfer knowledge from LW-DED (source) to LP-DED (target), thereby facilitating process map reconstruction of new materials. Here, synthetic SS316L LW-DED data and experimental IN718 LP-DED data are employed. To tackle the difference in parameter spaces that reflect distinct flow-thermal interactions in the two processes, input mapping calibration (IMC) is applied on Gaussian process (GP) regressors trained with source (LW-DED) data, yielding the first knowledge transfer model (IMC + GP). Then, it is integrated with a multi-fidelity GP (MFGP) framework to construct the second model (IMC + MFGP). Results show that when the target training set is small, both IMC + GP and IMC + MFGP outperform the baseline GP model. They can also achieve positive knowledge transfer with a small experimental source dataset. Furthermore, as compared with IMC + GP, IMC + MFGP better suits the scenarios when conservative uncertainty estimates are vital. Finally, both models are found to be sensitive to the variation of the IMC-related parameters due to the nominal relations between the input parameters. In conclusion, this research confirms the capability of transferring melt pool knowledge between LW-DED and LP-DED, and can potentially facilitate process map reconstruction of additively manufactured materials.