Abstract
The computational design of industrially relevant welded structures is extremely time consuming due to coupled physics and high nonlinearity. Previously, most welding distortion and residual stress simulations have been limited to small coupons and reduced order (from three-dimensional [3D] to two-dimensional [2D]), or inherent strain approximations were used for large structures. In this current study, an explicit finite element code based on a graphics processing unit (GPU) was utilized to perform 3D transient thermomechanical simulation of structural components during welding. Laser brazing of aluminum alloy panels and multipass arc welding of a nuclear industry pressurizer surge nozzle were simulated. The results demonstrate massive speedups—reducing simulation times from weeks to hours—while maintaining high fidelity in predicting residual stresses and distortions.
Introduction
Welding and brazing are fundamental manufacturing processes that introduce intense, localized heat. This thermal cycle causes structural distortion and "residual stresses"—internal stresses that remain after the component cools. Predicting these effects is critical for ensuring the structural integrity of safety-critical components, such as nuclear reactor nozzles, and the aesthetic quality of consumer products like automotive body panels.
However, high-fidelity 3D simulations of these processes are notoriously slow. Standard commercial software, typically relying on implicit Finite Element Method (FEM) solvers, often takes days or weeks to simulate even a single meter of weld. To manage this computational burden, engineers frequently resort to simplified 2D models or "inherent strain" approximations. While useful, these shortcuts often fail to capture critical 3D behaviors, such as complex clamping effects or localized stresses at weld starts and stops.
Methodology: GPU-Accelerated Explicit FEM
To overcome these limitations, this research shifts from the industry-standard implicit FEM toward a GPU-accelerated explicit FEM approach.
Parallel Computing Architecture
The core innovation lies in custom CUDA code designed to offload heavy computational tasks to thousands of GPU cores. Unlike standard implicit solvers that require solving massive simultaneous linear equations—a process where complexity grows exponentially ($O(N^2)$ to $O(N^3)$)—the explicit FEM scales linearly ($O(N^1)$). This linear scaling is essential for modeling large-scale industrial structures with millions of elements.
The workload is partitioned into specialized kernels: one for element-level stress/strain calculations (accounting for approximately 80% of processing time) and another for nodal time integration.
Time Scaling and Hybrid Workflow
To further accelerate the simulation, a "time scaling" factor is employed. This technique artificially shortens the simulation's time span without sacrificing physical accuracy, thereby reducing the total number of required integration steps.
The researchers also implemented a hybrid workflow. The fast explicit solver handles the complex, non-linear welding process, while a traditional implicit solver is used only for the final "clamping removal" stage, where the physics are simpler and converge rapidly.
Industrial Case Studies
The methodology was validated against two distinct industrial scenarios:
- Automotive Manufacturing: Laser brazing of AA6111 aluminum alloy panels (coach peel joint). Simulation results were compared against physical measurements from an ATOS optical 3D laser scanner.
- Nuclear Industry: The NRC/EPRI Phase IIb round robin mockup, involving a complex dissimilar metal weld (DMW) between a carbon steel nozzle and a stainless steel pipe. Validation was performed using experimental data from deep hole drilling.
Key Results and Discussion
The GPU-explicit method achieved unprecedented speedups:
- Automotive Panel (621k nodes): The simulation was completed in just 2 hours, compared to an estimated 252 hours (10.5 days) using standard commercial software on 8 CPU threads.
- Nuclear Nozzle (1.27M elements): The 3D simulation took 3.2 days. In contrast, standard software was estimated to require 19.6 years to complete the same task on 8 CPU threads.
Beyond speed, the 3D approach provided superior accuracy. In the nuclear case, the 3D model predicted residual stress distributions (hoop and axial) that matched experimental data significantly better than traditional 2D axisymmetric models. In the automotive case, the simulation captured complex "bulging" and "twisting" modes that are invisible to 2D approximations.
Conclusion
By leveraging GPU parallelization and explicit FEM, this study demonstrates that high-fidelity 3D welding simulations of large-scale structures are now feasible within industrial timeframes. This capability enables "what-if" scenarios for complex repairs and optimizes manufacturing processes to minimize distortion and residual stress, ultimately improving product safety and quality. Future work will focus on porting thermal conduction analysis to the GPU to eliminate remaining bottlenecks.
Summary updated