Numerical simulation of the collision dynamics in thin-walled structures under oblique load involves complex spatiotemporal processes, including material, geometric, and contact nonlinearities, which often require significant computational resources and time. Moreover, predicting high-dimensional spatiotemporal responses remains a challenge for most surrogate-based models. This paper proposes a deep learning framework based on manifold learning for spatiotemporal modeling of collision dynamics in thin-walled structures under oblique load. The framework leverages multiple deep learning models, including Variational Autoencoders (VAE), Radial Basis Function Interpolation (RBFI), and regression Residual Network (ResNet18), to capture the complex nonlinearities inherent in structural deformation, stress distribution, and crush force, enabling continuous prediction of multimodal spatiotemporal responses. Using a rectangular thin-walled tube under oblique load as an example, the models are validated with simulation data, yielding average prediction errors of 5.80 % for structural deformation, 6.01 % for Energy Absorption (EA), 10.66 % for Peak Crush Force (PCF), and 16.66 % for crush force. Compared to traditional finite element (FE) simulations, prediction time is reduced by 98.6 % for structural deformation and stress distribution, and 97.4 % for crush force. Additionally, the method demonstrates stability and broad applicability across different design parameters and structural configurations, including rectangular and double-cell tubes. This work underscores the potential of deep learning techniques to enhance computational efficiency and predictive accuracy in the crashworthiness design of thin-walled structures.

中文翻译：

---

基于流形学习的斜载荷下薄壁结构碰撞动力学预测的时空建模


斜载荷下薄壁结构碰撞动力学的数值模拟涉及复杂的时空过程，包括材料、几何和接触非线性，这通常需要大量的计算资源和时间。此外，预测高维时空响应仍然是大多数基于代理的模型的挑战。本文提出了一种基于流形学习的深度学习框架，用于斜载荷下薄壁结构碰撞动力学的时空建模。该框架利用多个深度学习模型，包括变分自动编码器 （VAE）、径向基函数插值 （RBFI） 和回归残差网络 （ResNet18），来捕获结构变形、应力分布和挤压力中固有的复杂非线性，从而能够连续预测多模态时空响应。以斜载荷下的矩形薄壁管为例，用仿真数据对模型进行了验证，结构变形的平均预测误差为 5.80 %，能量吸收 （EA） 为 6.01 %，峰值压碎力 （PCF） 为 10.66 %，压碎力为 16.66 %。与传统的有限元 （FE） 仿真相比，结构变形和应力分布的预测时间缩短了 98.6%，挤压力的预测时间缩短了 97.4%。此外，该方法在不同的设计参数和结构配置（包括矩形和双单元管）中表现出稳定性和广泛的适用性。这项工作强调了深度学习技术在提高薄壁结构耐撞性设计中的计算效率和预测准确性的潜力。