Physics-informed neural networks (PINN) can be used to predict flow fields with a minimum of simulated or measured training data. As most technical flows are turbulent, PINNs based on the Reynolds-averaged Navier–Stokes (RANS) equations incorporating a turbulence model are needed. Several studies demonstrated the capability of PINNs to solve the Naver–Stokes equations for laminar flows. However, little work has been published concerning the application of PINNs to solve the RANS equations for turbulent flows. This study applied a RANS-based PINN approach to a backward-facing step flow at a Reynolds number of 5100. The standard k-$\omega$ model, the mixing length model, an equation-free ${\nu }_t$ and an equation-free pseudo-Reynolds stress model were applied. The results compared favorably to DNS data when provided with three vertical lines of labeled training data. For five lines of training data, all models predicted the separated shear layer and the associated vortex more accurately.

Physics-informed neural networks are a promising method to yield surrogate models of flow fields. We present a metamodeling technique for variable geometries based on physics-informed neural networks. The method was applied to the DU99W350 airfoil at a Reynolds number of $$1\times 10^{5}$$1 × 10 5 . Using our technique, the angle of attack was introduced as an additional input parameter of the network and the model was trained to predict the Reynolds-averaged velocity and pressure fields around the airfoil for arbitrary angles of attack between 10.0 and 17.5 $$^{\circ }$$∘ . Furthermore, we present an effective method to generate the training points for the parameterized geometry. The model was trained with data from simulations for a limited set of angles of attack. Additionally, satisfaction of the a priori known boundary conditions as well as the Reynolds-averaged Navier–Stokes equations was attained. A sensitivity analysis concerning the Reynolds number, the amount and distribution of training data, and the turbulence model was conducted showing the superiority of the pseudo-Reynolds stress method and the demand for labeled training data in the domain. The trained network was capable of predicting the flow separation progressing with angle of attack on the suction surface and exhibited excellent agreement with numerically simulated results, even in the proximity of the wall for interpolations as well as extrapolations from the labeled data set. Our study demonstrates that physics-informed neural networks can be used to obtain accurate flow field surrogate models of variable geometries.

Unsupervised physics-informed deep learning can be used to solve computational physics problems by training neural networks to satisfy the underlying equations and boundary conditions without labeled data. Parameters such as network architecture and training method determine the training success. However, the best choice is unknown a priori as it is case specific. Here, we investigated network shapes, sizes, and types for unsupervised physics-informed deep learning of the two-dimensional Reynolds-averaged flow around cylinders. We trained mixed-variable networks and compared them to traditional models. Several network architectures with different shape factors and sizes were evaluated. The models were trained to solve the Reynolds-averaged Navier-Stokes equations incorporating Prandtl’s mixing length turbulence model. No training data were deployed to train the models. The superiority of the mixed-variable approach was confirmed for the investigated high Reynolds number flow. The mixed-variable models were sensitive to the network shape. For the two cylinders, differently deep networks showed superior performance. The best fitting models were able to capture important flow phenomena such as stagnation regions, boundary layers, flow separation, and recirculation. We also encountered difficulties when predicting high Reynolds number flows without training data.

Passive flow control techniques are needed to reduce flow separation and enhance aerodynamic performance. In this work, the effect of a knitted wire mesh on the flow separation of a backward-facing ramp was numerically investigated for a Reynolds number of 3000. A grid independence study and a RANS turbulence model sensitivity analysis were conducted. The CFD simulations exhibited counter-rotating streamwise vortices emerging from the knitted wire mesh, and the reattachment length was significantly reduced. A variation of the knitted wire rows revealed a maximum reduction of the reattachment length of 25.7% for the case of four rows. A comparison with a different knitted wire mesh geometry yielded a decreased reattachment length reduction.

Unsupervised physics-informed deep learning can be used to solve computational physics problems by training neural networks to satisfy the underlying equations and boundary conditions without labeled data. Parameters such as network architecture and training method determine the training success. However, the best choice is unknown a priori as it is case specific. Here, we investigated network shapes, sizes, and types for unsupervised physics-informed deep learning of the two-dimensional Reynolds-averaged flow around cylinders. We trained mixed-variable networks and compared them to traditional models. Several network architectures with different shape factors and sizes were evaluated. The models were trained to solve the Reynolds-averaged Navier-Stokes equations incorporating Prandtl's mixing length turbulence model. No training data were deployed to train the models. The superiority of the mixed-variable approach was confirmed for the investigated high Reynolds number flow. The mixed-variable models were sensitive to the network shape. For the two cylinders, differently deep networks showed superior performance. The best fitting models were able to capture important flow phenomena such as stagnation regions, boundary layers, flow separation, and recirculation. We also encountered difficulties when predicting high Reynolds number flows without training data

Physics-informed neural networks can be trained to serve as surrogate models by learning how spatial physical fields develop with change of parameters. This is an advantage over traditional numerical methods that can only be used to calculate a single solution field for a single discrete setting of a parameter. While many studies report accurate results of physics-informed deep learning of laminar flows, evaluations of surrogate modeling of high Reynolds number flows remain sparse. To contribute to this question, we here explore the capabilities and limits of physics-informed deep learning to solve the Reynolds-averaged Navier-Stokes equations for the flow around two airfoils under variable angles of attack. No labeled training data were provided and a single network was trained to simultaneously learn several solution fields in a single training run. The network captured essential flow features such as boundary layers and high and low pressure regions. Moreover, the qualitative correlations of lift and drag with the angle of attack and the higher drag of the thicker airfoil were captured. However, comparisons with reference simulations revealed an underestimation of the shear layer gradients as well as deviations of the fields' extrema. Thus, we also debate potential future improvements to increase the accuracy of the predicted flow. This work offers a prospect on the capabilities and current limitations of physics-informed neural networks for surrogate modeling of high Reynolds number flows. We argue that the method is promising and this work is aimed to encourage further research and development of the method when applied to Reynolds-averaged flows at high Reynolds numbers.