Embedded Files
Physics-informed neural networks (PINNs) for predicting flow fields
Physics-informed neural networks (PINNs) for predicting flow fields
It was investigated when physical constraints improve neural-network-based flow prediction compared with purely data-driven models. By varying the amount, distribution, and noise of training data, it was shown that physics-informed models are particularly beneficial when flow data are sparse, spatially constrained, or noisy, providing guidelines for efficiently combining physical knowledge with machine learning.
Physics-aware neural networks (PANNs) for initializing flow simulations
Physics-aware neural networks (PANNs) for initializing flow simulations
A data-free physics-aware convolutional neural network (PA-CNN) is being developed that predicts flow fields from geometry to provide improved initial conditions for high-fidelity CFD simulations. By learning directly from the Navier–Stokes equations rather than ground-truth CFD data, the approach accelerates convergence while preserving the accuracy of the final solution. Tested across 2D and 3D flows with complex geometries, the method reduced the required CFD iterations by up to 23% compared with conventional initialization strategies. The figure is extracted from [REFERENCE].
Page updated
Google Sites
Report abuse