Efficient prediction of thermal history and molten pool shape in a large domain for metal additive manufacturing via surrogate modeling and machine learning

This study is concerned with predicting accurate temperature fields in a large domain using a surrogate modeling technique and machine learning for metal laser powderbed fusion processes. Though high-fidelity modeling has been demonstrated to provide accurate representations of the attendant molten pool geometry and temperature field, simulation of large scale components is not feasible, even with massively parallelized computing, due to attendant prohibitive high computing costs. On the other hand, low fidelity models, albeit their computationally efficiency, cannot predict temperature history and molten pool shapes accurately and hence cannot be used for predicting resultant microstructure and process optimization. To overcome this challenge, a novel surrogate modeling approach is presented, which utilizes a lower-fidelity model to extract temperature variation over a large domain with pertinent features for use in a Gaussian process regression and implements an active learning algorithm to determine when and where the high-fidelity model, which has been developed in the authors’ group over the years, needs to be simulated to improve modeling results. Using such an approach, the overall computational load can be significantly decreased by a few orders of magnitude, thereby making it possible to accurately predict the temperature history, molten pool shapes and cooling rate. The surrogate modeling approach considers so called edge effects, i.e., distance from the edges, heat accumulation due to multiple scanning, and laser scanning patterns, as well as various laser parameters. Due to its computational efficiency, it is applicable to large parts with complex geometry.