Efficient GPU-accelerated thermomechanical solver for residual stress prediction in additive manufacturing

This paper addresses the need of rapid thermomechanical simulation of metal additive manufacturing by presenting a fully vectorized implementation of predicting the displacement field and residual stress for computation on graphical processing units. The formulation is based on implicit time discretization and the finite element method, where the incremental elastoplastic problem is solved using the conjugate gradient method at each Newton iteration. Sparse representation of algorithmic (tangent) stiffness matrix and the strain-displacement operator, are used in this formulation. A combined hardening plastic model, along with temperature-dependent material properties, is utilized. The temperature field is obtained by conducting detailed part-level thermal simulation using the explicit finite element method and then used in the mechanical simulation to calculate residual stresses. The details of the implementation of the proposed method are provided. Three simulation examples are performed to validate the thermomechanical model, compare the computational efficiency on GPU and CPU, and study the influence of toolpath strategy on residual stresses, respectively. In the example cases, the developed GPU implementation is 10-25x faster than the CPU version. The success of this development enables fast prediction of residual stress in additive manufacturing to improve the effectiveness of process design and to avoid process defects such as distortion and residual-stress induced fracture.

Automated summary

This paper presents a method for rapid thermomechanical simulation of metal additive manufacturing using a fully vectorized implementation on graphical processing units. The method utilizes implicit time discretization and the finite element method, solving the incremental elastoplastic problem using the conjugate gradient method. Sparse representation of algorithmic (tangent) stiffness matrix and the strain-displacement operator are used. The GPU implementation is 10-25x faster than the CPU version, enabling fast prediction of residual stress in additive manufacturing to improve process design effectiveness and avoid process defects.