Data-driven machine learning enhanced optimisation of vehicle crashworthiness design

耐撞性设计是汽车设计的一个重要方面。先进车辆安全性的进一步发展和轻量化结构的使用需要强大的优化策略。车辆碰撞模拟是计算密集型的，并且通常使用替代模型代替整车模型以降低模型的复杂性和计算工作量。虽然这些模型节省了时间和计算成本，但结果是次优的，并且在物理替代物的情况下，由于模型的简化，无法确定整体车辆响应。在这项研究中，我们利用机器学习（ML）方法来解决这些问题。强化学习（RL）是ML的一个子集，是一种强大的优化工具，但很少用于车辆设计。它有可能从经验中学习，并有可能产生接近最佳的参数。在这项研究中，提出了两种新的基于深度卷积生成对抗网络（DCGAN）的方法，一种基于软演员-评论家代理（SAC）的RL方法和两种监督学习神经网络（SLNN），以研究车辆耐撞性的多维优化。第一DCGAN用于生成合成数据，用于与仿真数据一起沿着训练第一SLNN，以提高训练精度。第二个SLNN被训练为连续体材料模型的数学代理，以加速FE模拟（申请人的专利）。基于SAC Agent的强化学习框架具有样本学习效率高、熵最大化能力强、稳定性好等特点，可用于汽车耐撞性优化设计。然后，第一个SLNN被用作所提出的基于深度SAC代理的RL网络的环境，该网络优化了设计参数。最后，第二DCGAN是用来估计减少代理模型的整体车辆响应。该研究旨在现代化和优化车辆耐撞性设计中的设计优化过程，与现有的替代模型相比，减少了总体计算工作量并提高了精度。

Design of Crashworthiness is a key aspect of vehicle design. The further evolution of advanced vehicle safety and usage of lightweight structures require powerful optimization strategies. Vehicle crash simulations are computationally intensive and often surrogate models are used in place of the full vehicle model to reduce the complexity of the model and computational effort. Although these models save time and computational cost, the results are sub-optimal and in the case of physical surrogates, the overall vehicle response cannot be determined due to the reduction of the model. In this study, we utilize machine learning (ML) methods to address these issues. Reinforcement learning (RL), which is a subset of ML, is a powerful optimization tool but has rarely been utilized in vehicle design. It has the potential to learn from experience and has the potential to generate near-optimal parameters. In this study, two novel Deep convolutional generative adversarial network (DCGAN) based approaches, an RL approach based on a soft actor-critic agent (SAC), and two supervised learning neural networks (SLNN) are proposed to investigate multidimensional optimization of crashworthiness of a vehicle. The first DCGAN is used to generate synthetic data for training the first SLNN along with simulation data to improve training accuracy. The second SLNN is trained as a mathematical surrogate for continuum material models to accelerate FE simulation (with a patent of the applicant). Due to its sample efficient learning and entropy maximization capability and stability, a SAC agent-based RL framework is used to optimize the vehicle crashworthiness design. The first SLNN is then used as the environment for the proposed deep SAC agent-based RL network which optimizes the design parameters. Finally, the second DCGAN is used to estimate overall vehicle response from reduced surrogate models. The study aims to modernize and optimize the design optimization process in vehicle crashworthiness design reducing overall computational effort and improving accuracy compared to existing surrogate models.