Search

Three reduced order models are evaluated in their capacity to predict the future state of an unsteady chaotic flow field. A spatially fully developed flow generated in a random packing of cylinders at a solid fraction of 0.1 and a nominal Reynolds number of 50 is investigated. For deep learning (DL), convolutional autoencoders are used to reduce the high-dimensional data to lower dimensional latent space representations of size 16, which were then used for training the temporal architectures. To predict the future states, two DL based methods, long short-term memory and temporal convolutional neural networks, are used and compared to the linear dynamic mode decomposition (DMD). The predictions are tested in their capability to predict the spatiotemporal variations of velocity and pressure, flow statistics such as root mean squared values, and the capability to predict fluid forces on the cylinders. Relative errors between 15% and 20% are evident in predicting instantaneous velocities, chiefly resulting from phase differences between predictions and ground truth. The spatial distribution of statistical second moments is predicted to be within a maximum of 5%–10% of the ground truth with mean error in the range of 1%–2%. Using the predicted fields, instantaneous fluid drag force predictions on individual particles exhibit a mean relative error within 20%, time-averaged drag force predictions to within 5%, and total drag force over all particles to within 1% of the ground truth values. It is found that overall, the non-linear DL models are more accurate than the linear DMD algorithm for the prediction of future states. [ABSTRACT FROM AUTHOR]

Solid-fluid multiphase systems are ubiquitous in many chemical, pharmaceutical, and energy based applications. These flows are challenging to study experimentally, thus several numerical simulation techniques have been developed. In this paper we use Particle Resolved Simulations (PRS) combined with Deep Learning (DL) to predict drag forces on each particle in suspensions of ellipsoidal particles for use in lower fidelity Euler-Lagrange and Euler-Euler simulations. Suspensions with solid fraction of 0.1, 0.2 and 0.3 are investigated at Reynolds numbers of 10, 50, 100 and 200. A UNet architecture is used to predict the velocity and pressure fields and a CNN is used to predict the drag forces. It is shown that drag force predictions using the predicted velocity and pressure fields give more accurate results than predictions that only use geometric information. [Display omitted] • Deep learning used to predict fluid drag forces in suspension of ellipsoids. • UNet architecture is used to predict interstitial flow and CNN to predict drag force. • Model accuracy evaluated for different input features and supervisions. • Flow predictions superior when informed by location of distant particles. • Drag force predictions are superior when informed by interstitial flow. [ABSTRACT FROM AUTHOR]

Two deep learning methods, Multi-Layer Perceptron (MLP) network and Convolution Neural Network (CNN) are evaluated to predict drag forces in dense suspensions of ellipsoidal particles using data from Particle Resolved Simulations (PRS). The MLP is trained on the mean flow Reynolds number, solid fraction of the suspension, the aspect ratio of the particle, and orientation to flow direction. The CNN is given an additional 3D spatial map of the particle of interest and its immediate neighborhood via a distance function. The prediction capability of the trained networks is tested at different levels of complexity: on an unseen particle arrangement (Level 1), to all arrangements of an unseen numerical experiment (Level 2), and finally to all experiments of an unseen Reynolds number, solid fraction or aspect ratio (Level 3). The CNN is shown to perform better than the MLP for all testing levels except when testing on an unseen aspect ratio. [Display omitted] • Fluid forces influence suspension dynamics in many engineering processes. • Deep learning is used to model drag forces on individual particles in suspension. • Model predictions are tested on different categories of unseen data. • Multi-layer perceptron and convolution neural network models are trained. • Convolutional Neural Networks are shown to perform better. [ABSTRACT FROM AUTHOR]