Exploration of Physics Informed Machine Learning Models for Turbulent Flows using DNS Data

Direct Numerical Simulation (DNS) and Large Eddy Simulation (LES) offer higher accuracy by resolving turbulence at finer scales but require significant computational resources and are impractical for real-world applications because of the high cost. An approach that would reduce computational time and deliver an accurate high-accuracy approximation is needed. To meet this need, high-quality DNS data for complex flows are used in this study to train a deep Neural Network based on relevant features of the flow fields, to develop regression models for Reynold stresses as a function of these features. We have validated this approach using DNS data for a single geometry with different Reynolds Number - flow over periodic hills at ���� = 5600 and ���� = 10595. The model’s predictive capability is assessed using data that have not been used in training, and comparing the predicted Reynolds stress to DNS based estimates.