Recent advancements in artificial intelligence (AI) have significantly influenced various fields, including mechanical engineering. Nonetheless, the development of high-quality, diverse datasets for structural analysis still needs to be improved. Although traditional datasets, such as simulated jet engine bracket dataset, are useful, they are constrained by a small number of samples, which must be improved for developing robust data-driven surrogate models. This study presents the DeepJEB dataset, which has been created using deep generative models and automated engineering simulation pipelines, to overcome these challenges. Moreover, this study provides comprehensive 3D geometries and their corresponding structural analysis data. Key experiments validated the effectiveness of the DeepJEB dataset, demonstrating significant improvements in the prediction accuracy and reliability of surrogate models trained on this data. The enhanced dataset showed a broader design space and better generalization capabilities than traditional datasets. These findings highlight the potential of DeepJEB as a benchmark dataset for developing reliable surrogate models in structural engineering. The DeepJEB dataset supports advanced modeling techniques, such as graph neural networks (GNNs) and high-dimensional convolutional networks (CNNs), leveraging node-level field data for precise predictions. This dataset is set to drive innovation in engineering design applications, enabling more accurate and efficient structural performance predictions. The DeepJEB dataset is publicly accessible at: https://www.narnia.ai/dataset
翻译:近年来,人工智能(AI)的显著进展对包括机械工程在内的多个领域产生了深远影响。然而,面向结构分析的高质量、多样化数据集的开发仍有待加强。尽管传统数据集(如模拟喷气发动机支架数据集)具有实用价值,但其受限于样本数量较少,这对于开发稳健的数据驱动代理模型而言亟需改进。本研究提出的DeepJEB数据集,通过深度生成模型与自动化工程仿真流程构建,旨在应对上述挑战。此外,本研究提供了完整的三维几何结构及其对应的结构分析数据。关键实验验证了DeepJEB数据集的有效性,表明基于该数据训练的代理模型在预测精度与可靠性方面均有显著提升。相较于传统数据集,增强后的数据集展现出更广阔的设计空间与更优的泛化能力。这些发现凸显了DeepJEB作为结构工程中开发可靠代理模型的基准数据集的潜力。DeepJEB数据集支持图神经网络(GNNs)与高维卷积网络(CNNs)等先进建模技术,可利用节点级场数据实现精确预测。该数据集有望推动工程设计应用的创新,实现更准确高效的结构性能预测。DeepJEB数据集已公开访问,地址为:https://www.narnia.ai/dataset