Researchers Rong Ma and Adrian Lozano-Duran from the Massachusetts Institute of Technology (MIT) have developed a novel approach to improve the accuracy of large-eddy simulations (LES) in predicting fluid flow and heat transfer over rough surfaces. Their work, published in the Journal of Fluid Mechanics, presents a machine-learning-based wall model that can be applied across a wide range of flow speeds and surface roughness conditions.
Large-eddy simulations are a powerful tool in computational fluid dynamics, allowing researchers to model turbulent flows by simulating the larger eddies and modeling the smaller ones. However, accurately predicting the effects of surface roughness on flow and heat transfer has been a persistent challenge. Ma and Lozano-Duran’s new model addresses this issue by incorporating surface-roughness effects into LES, making it applicable to both low- and high-speed flows, as well as transitional and fully rough conditions.
The researchers trained their model using an artificial neural network on a comprehensive database of direct numerical simulations (DNS) of compressible turbulent channel flows over rough walls. The dataset included 372 cases with varying roughness topographies, Mach numbers ranging from 0 to 3.3, and friction Reynolds numbers between 180 and 2000. To ensure the model’s predictive power, they employed an information-theoretic, dimensionless learning method to identify the most relevant inputs for predicting dimensionless wall friction and wall heat flux.
One of the key features of the new model is its ability to quantify uncertainty. By using a spectrally normalized neural Gaussian process, the model provides a confidence score that indicates the level of uncertainty in regions that deviate from the training dataset. This feature is particularly valuable in the energy sector, where accurate predictions of flow and heat transfer are crucial for designing and optimizing energy systems.
The researchers evaluated the model’s performance both a-priori and a-posteriori. A-priori tests on 110 turbulent channel flow cases yielded prediction errors below 4%. A-posteriori assessments in wall-modeled LES across diverse test cases, including subsonic and supersonic turbulent channel flows, a transonic high-pressure turbine blade, a high-speed compression ramp, and hypersonic blunt bodies, demonstrated the model’s versatility and accuracy. The model typically achieved predictive accuracy within 10% for wall shear stress and within 15% for wall heat flux, with high confidence in most cases. In scenarios with the highest errors, the model correctly detected the reduced performance through a drop in the confidence score.
The practical applications of this research for the energy sector are significant. Accurate predictions of flow and heat transfer over rough surfaces are essential for designing and optimizing various energy systems, such as turbines, compressors, and heat exchangers. By improving the accuracy of LES, the new wall model can help engineers develop more efficient and reliable energy technologies, ultimately contributing to a more sustainable energy future.
Source: Ma, R., & Lozano-Duran, A. (2023). Machine-learning wall model of large-eddy simulation for low- and high-speed flows over rough surfaces. Journal of Fluid Mechanics.
This article is based on research available at arXiv.