Computational Fluid Dynamics (CFD) simulation of viscoplastic fluid flows is challenging. It is critical to capture the shape of the yield surfaces that limit yielded and unyielded flow regions. These are highly sensitive to numerical schemes, which usually rely on calculating velocity derivatives and require highly refined computational grids and strict convergence criteria. In this work, CFD and Deep Neural Networks (DNN) were coupled to develop a model to predict yield surface morphologies for the flow of Bingham fluids in a square cavity. Our main goal is to predict the behavior of viscoplastic fluids using a Machine Learning (ML) approach based on CFD results. Design of Experiments techniques were employed in defining the base cases, i.e., combinations of Bingham and Reynolds numbers (Bn and Re) to compose the DNN training data. The DNN predicted the viscosity fields in the problem domain given Bn and Re. These results were postprocessed using masks to create binary images. The chosen architecture was an encoder-decoder since the input and output data had different dimensions. The results of the surrogate model were adequate, giving a Mean Squared Error of 0.0015 for the training data and 0.002 for the testing data. The DNN-predicted images were consistent with those generated from CFD, corroborating the proposed technique as an excellent alternative to be implemented in more complex applications. The combination of CFD and ML is a promising alternative for predicting complex fluid behavior in diverse and challenging scenarios with faster and computationally less expensive resources.
