Key Laboratory for Mechanics in Fluid Solid Coupling Systems, Institute of Mechanics, Chinese Academy of Sciences, Beijing, 100190, China.
School of Engineering Science, University of Chinese Academy of Sciences, Beijing, 100049, China.
Sci Rep. 2020 Mar 10;10(1):4459. doi: 10.1038/s41598-020-61450-z.
In a myriad of engineering situations, we often hope to establish a model which can acquire load conditions around structures through flow features detection. A data-driven method is developed to predict the pressure on a cylinder from velocity distributions in its wake flow. The proposed deep learning neural network is constituted with convolutional layers and fully-connected layers: The convolutional layers can process the velocity information by features extraction, which are gathered by the fully-connected layers to obtain the pressure coefficients. By comparing the output data of the typical network with Computational Fluid Dynamics (CFD) results as reference values, it suggests that the present convolutional neural network (CNN) is able to predict the pressure coefficient in the vicinity of the trained Reynolds numbers with various inlet flow profiles and achieves a high overall precision. Moreover, a transfer learning approach is adopted to preserve the feature detection ability by keeping the parameters in the convolutional layers unchanged while shifting parameters in the fully-connected layers. Further results show that this transfer learning network has nearly the same precision while significantly lower cost. The active prospects of convolutional neural network in fluid mechanics have also been demonstrated, which can inspire more kinds of loads prediction in the future.
在众多工程情况下，我们通常希望建立一个模型，通过检测流场特征来获取结构周围的载荷情况。本文提出了一种数据驱动的方法，通过对圆柱尾迹流场的速度分布进行预测，从而获得圆柱表面的压力系数。所提出的深度神经网络由卷积层和全连接层组成：卷积层可以通过特征提取来处理速度信息，然后通过全连接层将这些特征汇集起来，以获得压力系数。通过将典型网络的输出数据与计算流体动力学（CFD）结果进行比较，结果表明，该卷积神经网络（CNN）能够在各种入口流场条件下预测训练雷诺数附近的压力系数，并且具有较高的整体精度。此外，还采用了迁移学习方法，通过保持卷积层中的参数不变，同时调整全连接层中的参数，来保持特征检测能力。进一步的结果表明，这种迁移学习网络具有几乎相同的精度，而成本却显著降低。本文还展示了卷积神经网络在流体力学中的应用前景，这为未来的各种载荷预测提供了启示。
