Predicting Roughness Effects on Velocity and Temperature in Turbulent Flow - A Data-Driven Approach

Accurately predicting the impact of arbitrary rough surfaces on turbulent fluid flow is crucial for industrial applications and for enhancing the precision of RANS simulations. Thermohydraulic turbulent roughness properties, specifically the augmentation in velocity $\mathrm{\Delta }{U}^{+}$ and temperature $\mathrm{\Delta }{\mathrm{\Theta }}^{+}$, describe this effect, but their prediction for a rough surface currently requires a detailed simulation. Recently, the application of neural networks for these types of predictions have been developed to circumvent the costly simulations. $93$ high fidelity direct numerical simulations of a fully-developed turbulent channel flow at $R{e}_{\tau }\approx 800$ with artificially generated realistic rough surfaces provide a sufficient database for training, validation, and testing of a neural network architecture with strong predictive capabilities for $\mathrm{\Delta }{U}^{+}$. For better explainability of the 'black-box' model, a genetic programming technique, namely symbolic regression, is applied to translate the data-driven model into an understandable, simple expressions using well-known statistical parameters. Nevertheless, the prediction of the temperature augmentation using an analog procedure remains insufficient, particularly since the dependencies on the Prandtl number $Pr$ have not yet fully been investigated.
... mehr

As supercomputing facilities are transitioning towards heterogeneous cluster architectures and GPU accelerated computation showed massive speed-up in other domains, using these new capabilities to overcome the data limitation motivated the use of a GPU accelerated DNS-solver. Using this GPU-code for rough surface channel flow with non-conjugate heat transfer is tested, and the applicability for data generation on the Tier 2 Cluster HoreKa will be additionally discussed in the conference presentation.