Predicting residual deformation in LPBF with neural networks and uncertainty quantification

Principal Investigators: Zhang, Jessica, Levent Burak Kara

University: Carnegie Mellon University

Industry Partners: HexSpline3D LLC

Metal additive manufacturing (AM) has become popular for fabricating industrial parts. However, build failures due to geometry compensation constitutes a key impediment holding back the widespread adoption of metal AM. We propose a data-driven surrogate modeling approach to replace expensive numerical simulations of the laser powder bed fusion (LPBF) process. By speeding up geometry distortion calculations from several hours to mere seconds, our model can be deployed in structural design pipelines to prevent generation of infeasible designs. The proposal is a partnership between  researchers at Carnegie Mellon University and  a local PA startup company, HexSplines3D LLC. We will provide HexSplines3D with a competitive technological edge in (1) ML based implicit geometry representations, (2) data-driven modeling of part deformation in LPBF, and (3) an associated uncertainty quantification methodology, absent from existing solvers. Furthermore, we will generate a training dataset of residual deformation calculations that will be openly available to the AM community.