| Abstract Scope |
Metal Additive Manufacturing (AM) allows for great versatility in part fabrication but suffers from significant residual stress and distortion due to the extreme thermal gradients inherent to the melting process. While thermo-mechanical finite element analysis provides high-fidelity predictions, its computational cost is prohibitive for industrial-scale components. The Inherent Strain Method offers a rapid alternative to predict welding-induced distortion by treating it as a result of incompatible strains. More recently, the Modified Inherent Strain Method (MISM) refines this approach by specifically accounting for the global evolution of the elastic strain field under the evolving boundary conditions characteristic of the layer-wise AM process. However, the accuracy of MISM remains dependent on an Inherent Strain Parameter (ISP) that typically requires expensive experimental calibration or thermo-mechanical simulation of a representative sample part for every new material-parameter combination.
This work proposes a machine learning framework to predict the ISP within a bulk region of deposited material by leveraging latent vector representations of material properties and laser powder bed fusion (LPBF) processing parameters. By integrating a dimensional bottleneck into a Deep Kernel Learning (DKL) architecture, the model is forced to capture the underlying physical relationships governing distortion. The uncertainty quantification provided by DKL is then leveraged alongside thermo-mechanical simulations in an active learning loop to autonomously identify the next best sampling points in the parameter space, reducing the required training data. The model was validated against the distortion of an austenitic stainless steel 316L overhangs printed by different laser processing parameters. The developed model enables rapid, confidence-informed ISP estimation, facilitating accelerated distortion prediction for large-scale AM parts. |