Accelerated and Reduced-Order Simulations of Metal Additive Manufacturing Processes

Additive manufacturing (AM) techniques like laser powder-bed fusion and directed energy deposition allow efficientfabrication of complex metal parts through layer-by-layeraddition of material. Although the economic impact of AM isprojected to potentially exceed hundreds of billions ofdollars, its widespread adoption has been slowed in part byvariability in final material quality. Computational modelsare crucial in understanding the complex relationshipsbetween process, microstructure, and properties that affectmaterial quality and performance. Simulation techniqueshave been developed to successfully predict theserelationships in detail, but computational times are oftentoo slow to allow the rapid and iterative calculations neededfor design, optimization, and control of AM processes andparts. In this talk, I present advances in computationaltechniques that use a collection of physics-based anddata-driven techniques to accelerate AM processsimulation. In one approach, a novel scale separation is usedto allow efficient GPU implementation of temperaturehistory during an AM process. To reduce computationaltimes even further, we introduce reduced-order models forthe smallest scales using machine learning to identify andupdate a smaller set of latent variables. Finally, I show how areduced representation of the crystal grain structureenables effective calibration and validation of process-structure models for AM materials.