Presentation: GPU-accelerated optimization with structural analysis
Dr.-Ing. Daniel Weber, Session 5F - Solvers, 17th May 2023, 14:10
Finite element simulations are applied in a vast number of engineering tasks. For example, in virtual prototyping, the product development process, topology and shape optimization, material distribution optimization, physical data synthesis for machine learning, and additive manufacturing process simulation.
At their core, all these tasks require simulation of the physical behavior of a part. For structural analysis, the method of choice is the finite element method, which is computationally intensive. Minimizing the runtime for finite element simulations has a large impact on these tasks. Therefore, we develop massively parallel algorithms and dedicated data structures for our project Rapid Interactive Structural Analysis (RISTRA) to achieve a very high degree of parallelism and efficiency on graphics processing units (GPUs), which are also used outside of computer graphics, due to their high potential for computationally intensive tasks.
To achieve minimal run times for a single structural simulation, several aspects must be considered. The computation of the element stiffness matrices, the setup of the linear system, its solution and the results evaluation, all must be conducted efficiently. Additionally, transfer between GPU and CPU memory must be minimized, which we achieve by implementing all building blocks on the GPU. However, not only the computationally intensive tasks are of importance, but focus also needs to be put on efficient integration with software tools that use RISTRA. Many software toolchains still rely on inefficient plain-text file exchange. Using binary file exchange—or even better, setting up application program interfaces that enable a direct data transfer—drastically reduces overhead.
Taking all these aspects into consideration, the time for single simulation runs can be reduced drastically. Scenarios where tens or hundreds of simulation runs are necessary benefit immensely from the reduced runtime. Especially optimization tasks such as topology and shape optimization or computing optimal material distributions benefit greatly from such short simulation times. In this work, we investigate how optimization can benefit from GPU-accelerated simulation. We evaluate the runtimes for individual runs in comparison to commercially available software. Furthermore, we analyze how optimization scenarios benefit from a GPU-accelerated solver and compare result quality. Furthermore, we demonstrate in-situ visualization. As the solver is running on the GPU, we are able directly use its output for visualization without data transfer. Therefore, direct inspection of the results is possible, i.e., in-situ visualization of simulation and optimization is feasible. This enables users to cancel simulation runs if a configuration has been specified incorrectly or gather a deeper understanding of optimization algorithms.