Application of fully convolutional neural networks for feature extraction in fluid flow

Accurate extraction of features in fluid flows is of importance due to the presence in many natural and technological systems. Recently, methods based on machine learning have emerged as an alternative to traditional Eulerian-based methods to extract features in fluid flows. One broad category in ML is the convolution operation-based methods. The precision of feature extraction in convolution operation-based methods increases by constraining the measurement box, such as dividing the input data into small patches and using them as input boxes in convolutional neural networks. In this work, we propose a method that transforms each cell of the computational domain into a detection pixel (measurement box) to perform the task of feature extraction at the smallest possible computational level. To demonstrate the performance, we extract the vortical structures in a benchmark two-dimensional lid-driven cavity flow employing a symmetric, fully convolutional network. The number of convolution and deconvolution blocks in the network’s structure is studied to obtain the highest accuracy and yet to avoid the degradation problem. Different parameters, such as the Reynolds number and velocity boundary values, are considered in the complete and clipped cavity cases to create the training and test datasets. The semantic segmentation metrics, including Jaccard and Dice, yield values close to 1 for the test set on complete and clipped cavity cases with varying the Reynolds number or the velocity boundary values.