基于人工智能技术的钝体尾流时空预报

Spatiotemporal Forecast of Bluff‑body Wakes Using Artificial Intelligence Technologies

摘要:

湍流在机械工程、流体力学、土木工程等领域中广泛存在。以往，湍流信息获取主要依赖数值模拟和风洞试验，但数值模拟花费时间长，风洞试验经济成本高昂。新时代技术发展背景下，人工智能技术因其高效、高精度和可信赖等特点受到工程领域广泛关注。本研究构建了 Turbulent Flow?Vision?Transformer（TF?ViT）的人工智能算法，TF?ViT 可以基于数据驱动的方式来实现时空序列的湍流预报。具体来说，TF?ViT 主要包含两个部分：Transformer 框架及 UNet 结构。在 TF?ViT 算法中，每一个部分都有独特的功能，其中 Transformer 框架为编码器，主要用于湍流时空特征处理，UNet 网络是解码器，对编码器处理的时空湍流信息进行解耦，整体框架可以预测未来的时空湍流信息。用经典的矩形柱体绕流来验证开发的 TF?ViT 算法，开源计算求解器 OpenFOAM 用于矩形柱体绕流模拟，矩形柱体的尾流场数据用于 TF?ViT 模型的训练和验证。使用 8 帧连续瞬态湍流信息预报未来 8 帧的湍流信息，研究结果表明本研究开发的 TF?ViT 算法可以较为准确地预报未来短时间内尾流区的湍流时空发展。本研究展示了 TF?ViT 预报时空湍流的能力，为获得湍流尾流场提供了一种有效的方法。

Abstract:

Turbulent flow is ubiquitous in mechanical engineering, fluid mechanics, civil engineering, and other related disciplines. Traditionally, acquisition of turbulent flow data mainly depended on numerical simulations and wind tunnel tests. However, numerical simulations require substantial computational time, and wind tunnel tests involve high economic costs. With the rapid development of modern technologies, artificial intelligence technologies have attracted widespread attention in engineering fields due to their high efficiency, high precision, and reliability. This study developed an artificial intelligence algorithm named Turbulent-Flow-Vision Transformer (TF-ViT), which enabled spatiotemporal forecast of turbulent flow based on data-driven approaches. Specifically, the TF-ViT mainly consisted of two components: Transformer framework and UNet structure. In TF-ViT, each component had distinct functions. The Transformer framework served as the encoder, mainly responsible for processing and extracting spatiotemporal features of turbulent flow. Meanwhile, the UNet functioned as the decoder to decouple the encoded spatiotemporal turbulent flow information. The overall framework enabled the forecast of future spatiotemporal turbulent flow information. This study used the classical problem of the flow past rectangular cylinders to validate the developed TF-ViT algorithm. The open-source computational solver OpenFOAM was utilized to simulate the flow past rectangular cylinders, and the obtained wake flow field data was then used for the training and validation of the TF-ViT model. 8 continuous frames of transient turbulent flow data were used to forecast the subsequent 8 frames of turbulent flow information. The results showed that the developed TF-ViT algorithm in this study could accurately forecast the short-term spatiotemporal development of turbulent flow in the wake region. This study demonstrates the strong capability of TF-ViT in forecasting spatiotemporal turbulent flow, providing an effective method for turbulent wake field acquisition.