Turbulent flow over a series of increasingly high, two-dimensional bumps is studied by well-resolved large-eddy simulation. The mean flow and Reynolds stresses for the lowest bump are in good agreement with experimental data. The flow encounters a favourable pressure gradient over the windward side of the bump, but does not relaminarize, as is evident from near-wall fluctuations. A patch of high turbulent kinetic energy forms in the lee of the bump and extends into the wake. It originates near the surface, before flow separation, and has a significant influence on flow development. The highest bumps create a small separation bubble, whereas flow over the lowest bump does not separate. The log law is absent over the entire bump, evidencing strong disequilibrium. This dataset was created for data-driven modelling. An optimization method is used to extract fields of variables that are used in turbulence closure models. From this, it is shown how these models fail to correctly predict the behaviour of these variables near to the surface. The discrepancies extend further away from the wall in the adverse pressure gradient and recovery regions than in the favourable pressure gradient region.

No abstract

Turbulent flows involving adverse pressure gradients, curvature and mild separation are analyzed and data-driven augmentations are developed for predictive models. Large eddy simulations are performed over a set of parametric flow configurations and the impact of varying curvature, boundary layer thickness and Reynolds number is assessed. The first step in the data-driven methodology involves the determination of functional discrepancies in existing models using inverse modeling. The inferred discrepancy is reconstructedas a function of locally non-dimensional flow features-using a machine learning algorithm. This machinelearned discrepancy is embedded within a k − ω turbulence model. The impact of the choice of data used for the inverse modeling on the predicted velocity field and Reynolds stresses is analyzed in detail. The dataaugmented turbulence model, trained using a very small subset of flows and limited data is shown to yield much-improved predictions of flow properties, over the entire set of configurations. This work represents a key step toward the development of more general data-augmented turbulence models.

The performance of an industrial fan was simulated using CFD and results were compared with the experimental data. The fan is used to cool a row of resistor networks which dissipate excess energy generated by regenerative power in an inverter application. It has a diameter of 24 inches (0.6096m) and rotates at different speeds ranging from 2500 to 3900 RPM depending on the requirements. CFD simulation results were also verified by simulating performance of the same fan at different speeds and comparing the results with what was expected from fan affinity laws. The CFD results matched almost exactly (with ∼0.2% difference for pressure at a given flow rate) with the performance being predicted by the affinity laws. The effect of variation of different parameters such as the blade length, number of blades, and blade chord length was studied. Increasing the blade length at the same RPM increased the mass flow rate (by ∼17%) for the same pressure. Increasing the chord length while keeping the same number of blades, at a given RPM, made the performance curve (pressure versus flow rate, i.e. PV curve) steeper and blades stalled at a higher mass flow rate (8.77 kg/sec compared to the previous 8.44 kg/sec). For the same total blade surface area, less number of blades with longer chords stalled at lower mass flow rates (9.22 kg/sec for a 33% shorter chord and 36 blades compared to 8.3 kg/sec for the original rotor which had 24 blades).

Naïve estimation of horizontal wind velocity over complex terrain using measurements from a single wind-LiDAR introduces a bias due to the assumption of uniform velocity in any horizontal plane. While Computational Fluid Dynamics (CFD)-based methods have been proposed for bias removal, there are several issues exist in the implantation. For instance, the upstream atmospheric boundary layer thickness or direction are unknown. Conventional CFD-based corrections use trial and error to estimate the bias. Such approaches not only become numerically intractable for complicated flows, e.g. when the number of unknowns is large, but they also suffer from the fact that there is no guarantee for optimality of the obtained results. We propose a direct-adjoint-loop (DAL) optimization based framework to estimate such unknown parameters in a systematic way. For the validation of the method, we performed an experimental study using DIABREZZA LiDAR on a complex terrain for two wind directions of northwesterly (NW) and southeasterly (SE). The slope error associated with linear regression improved from -0.09 to -0.02 for SE and from -0.09 to +0.04 for NW.