No abstract

Feature identification is an important task in many fluid dynamics applications and diverse methods have been developed for this purpose. These methods are based on a physical understanding of the underlying behavior of the flow in the vicinity of the feature. Particularly, they rely on definition of suitable criteria (i.e. point-based or neighborhood-based derived properties) and proper selection of thresholds. For instance, among other techniques, vortex identification can be done through computing the Q-criterion or by considering the center of looping streamlines. However, these methods rely on creative visualization of physical idiosyncrasies of specific features and flow regimes, making them non-universal and requiring significant effort to develop. Here we present a physics-based, data-driven method capable of identifying any flow feature it is trained to. We use convolutional neural networks, a machine learning approach developed for image recognition, and adapt it to the problem of identifying flow features. The method was tested using mean flow fields from numerical simulations, where the recirculation region and boundary layer were identified in a two-dimensional flow through a convergent-divergent channel, and the horseshoe vortex was identified in three-dimensional flow over a wing-body junction. The novelty of the method is its ability to identify any type of feature, even distinguish between similar ones, without the need to explicitly define the physics (i.e. through development of suitable criterion and tunning of threshold). This provides a general method and removes the large burden placed on identifying new features. We expect this method can supplement existing techniques and allow for more automatic and discerning feature detection. The method can be easily extended to time-dependent flows, where it could be particularly impactful. For instance, it could be used in the identification of coherent structures in turbulent flows, a hindrance in the ongoing effort to establish a link between coherent structures and turbulence statistics.

No abstract

No abstract

Training data-driven turbulence models with high fidelity Reynolds stress can be impractical and recently such models have been trained with velocity and pressure measurements. For gradient-based optimization, such as training deep learning models, this requires evaluating the sensitivities of the RANS equations. This paper explores the use of an ensemble approximation of the sensitivities of the RANS equations in training data-driven turbulence models with indirect observations. A deep neural network representing the turbulence model is trained using the network's gradients obtained by backpropagation and the ensemble approximation of the RANS sensitivities. Different ensemble approximations are explored and a method based on explicit projection onto the sample space is presented. As validation, the gradient approximations from the different methods are compared to that from the continuous adjoint equations. The ensemble approximation is then used to learn different turbulence models from velocity observations. In all cases, the learned model predicts improved velocities. However, it was observed that once the sensitivity of the velocity to the underlying model becomes small, the approximate nature of the ensemble gradient hinders further optimization of the underlying model. The benefits and limitations of the ensemble gradient approximation are discussed, in particular as compared to the adjoint equations.