Efficient part-scale thermal modeling of laser powder bed fusion via a multilevel finite element framework

Abstract

In this work, we show that a multilevel finite element algorithm previously demonstrated for linear problems can, using a novel time integration method and other improvements, give efficient and accurate part-scale simulations of real additive manufacturing processes. The GPU-optimized multilevel finite element framework (GO-MELT) uses multiple moving meshes to simulate thermal behavior in laser powder bed fusion (LPBF) processes; fixed mesh sizes and data structures allow straightforward implementation of this algorithm on GPU hardware. Building on this framework, we introduce key advancements including G-code parsing for complex laser paths, temperature-dependent material properties with distinct definitions for powder, solid, and fluid states, and time step subcycling across levels to manage computational loads effectively. These improvements enable precise simulation across the different material states encountered in LPBF while minimizing computational cost. Verification studies show that first-order time convergence is preserved even in the presence of nonlinearities, and the fidelity of the enhanced framework is validated against well-established experimental benchmarks, including in-situ X-ray diffraction data for Hastelloy-X and time above melting measurements from the NIST AM-Bench cantilever model. Computational tests demonstrate that our approach achieves an average execution time of 1.8 ms per time step, enabling a high-fidelity thermal simulation of 350 million time steps to be solved on a single GPU in 7.3 days, comparable to published simulations on much larger parallel systems. An analysis of thermal decay times can be used to further reduce simulation time by limiting simulation to time-points of interest. These results underscore the potential of this algorithm for advancing real-time process optimization and part quality improvement in LPBF.