GPU And Reanalysis-based Efficient Optimization Method For Design Problems Of Automotive Body

Numerical simulations are commonly performed in the procedure of optimization design. However, with the increasing complexity and the scale of the engineering problem, the computational cost becomes large and it is urgent to overcome the bottleneck of low efficiency. On the one hand, fast computational methods such as reanalysis method can be employed to avoid complete analysis of the modified structure by using the result information of the initial structure. On the other hand, the most direct way of improving the efficiency is parallelsim. Numerous parallel applications have been implemented on the supercomputers as well as CPU-based clusters, but these platforms are extremely expensive and requries complex programming by the usage of coarse-grained parallel domain decomposition strategy. Furthermore, the computational efficiency is directly related to the number of computer nodes, making them less effective and accessible. Modern graphics processor unit(GPU) has become a multi-core processor with highly internal parallelism, and its peak computing power is orders of magnitude greater than that of modern CPUs, and 20 times or more higher memory bandwideth than CPU to memory interface. Nowadays, as a novel and intuitive tool for GPGPU, compute unified device architecture(CUDA) developed by NVIDIA makes it easy and efficient to program on GPU to attain high computing efficiency.In this paper, for the purpose of overcome the bottleneck of computational efficiency, a GPU-based parallel reanalysis system for vehicle structure design is presented to meet the demands of engineering applications. And, based on the proposed parallel reanalysis method, a parallel topology optimization method on normal personal computer with a CUDA-capable device is realized, which can accelerate the conceptual design cycle of vehicle structures. Furthermore, in order to accelerate the computational efficiency of detailed design cycle, an efficient global optimization method based on multiple data is proposed and implemented parallelly. So the parallel optimization method can be easily implemented based on normal computer. In order to overcome the performance bottleneck of single GPU, a multiple-GPU reanalysis system is established. The main research content and result are shown as follows:(1) A GPU-based parallel reanalysis platform based on the characteristics of reanalysis scheme and lightweight threads parallel computing model of GPU is presented. The main advantage of this platform is the construction of parallel computation and assembly for stiffness matrix based on two mapping strategies, parallel computation for inverse matrix, and parallel preconditioned conjugate gradient solver based on SSOR. Therefore, it is convenient to perform parallel reanalysis for large scale vehicle body structure with high efficiency.(2) This paper proposes a parallel topology optimization method based on GPU parallel reanalysis method. By employing thread mapping element strategy, parallel computation for sensitivity numbers is implemented on GPU. The parallel filtering for sensitivity numbers is also presented in order to enhance computation efficiency. By defining and evaluating the change of the structure, the parallel reanalysis is introduced and used when the change is small in which case the change is less than a tolerance value. This strategy makes good balance between computaion efficiency and computational accuracy.(3) This paper also presents a parallel efficient global optimization method based on multiple data to accelerate the optimization process. Instead of using multiple surrogates during the optimization procedure, this paper proposes an alternative strategy by emloying the multiple data generated by data classification. This method, named MDEGO, has shown better convegence compared with the popular MSEGO method with the same simulation numbers. And its parallel implementation based on GPU is implemented to enhance its efficiency for engineering problems.(4) The parallel platform based on single GPU has low efficiency and memory bottleneck. To breakthrough these bottlenecks, an efficient parallel reanalysis method is developed based on multiple GPUs platform. Data distribution is successfully implemented based on matrix-vertor decomposition, a parallel pre-conditioned conjugate gradient method is proposed based on multiple GPUs, and the reanalysis method is reconstructed for muliple GPUs. This paper also proposes an effective technique to overlap the computation and communication by using non-blocking communication strategy. Finally, large scale vehicle design problems demonstrate the effectiveness of the method. Compared with single GPU, this platform shows high speedup as well as the same accuracy.