A data-driven method is developed in this article to predict the pressure coefficients from the velocity distribution in the wake flow. The convolutional layer processes velocity information in local region to output flow feature, which are gathered by the fully connected layer to obtain the pressure coefficients. When meeting different around body flow situation, a transfer learning method is adopted. Results show that this transfer learning method achieves nearly the same accuracy as the traditional one but with significantly lower time cost. The learning results have also demonstrated the active prospects of convolutional neural network in fluid mechanics.
Haven gone through two ups and downs, machine learning gained its popularity in 2006 and lasts for a long time. Meanwhile, the emerge of big data era and the significantly improved computing ability of the devices have laid a good foundation for the machine learning, as well as provided a broad platform for related application and promotion. This application includes computer vision, speech recognition, recommender system, natural language processing (Lecun, 2015). However, many machine learning methods achieve good results relying on the training data and the testing data sharing the same feature space and the same distribution. For those cases whose wanted training data is impossible or difficult to collect, transfer learning between task domains is beneficial. Web-document classification, sentiment classification and indoor WiFi localization problems are examples of transfer learning helps in training the model (Pan, 2010). Today, transfer learning methods applied in many machine learning and data mining applications. Before we discuss the chance of apply it to the fluid mechanics, the current research status of machine learning in fluid mechanics is discussed as follows.
It was since a few years ago, when machine learning gains its popularity in many engineering field, the chances of applying it to fluid mechanics began to appear. Early attempt was made by Tracey (2013), who used neural networks with a single hidden layer to simulate terms in Spalart-Allmaras RANS model. This study suggested the feasibility of neural networks for turbulence modeling. Julia Ling (2016) proposed a multiplicative layer embedded with Galilean invariance to predict the Reynolds stress anisotropy tensor, and this neural network architecture provides improved accuracy than before. Wang (2017) reconstruct the discrepancies in the RANS-simulated Reynolds stress using high-fidelity DNS data based on machine learning methods, which offer a different way to interpret the improved prediction. However, more study needs to be done to figure out the chances of machine learning owning in fluid mechanics beyond turbulence modeling. Ströfer (2018) successfully identify the flow features and even distinguish similar ones by convolutional neural networks, a machine learning method for image recognition. Jin (2018) build a CNN architecture that can capture the spatial-temporal information and the translations features of the pressure fluctuation on the cylinder.
