Generalizability evaluation of k-ε models calibrated by using ensemble Kalman filtering for urban airflow and airborne contaminant dispersion

The low accuracy of Reynolds-averaged Navier-Stokes (RANS) modeling of urban airflow and airborne pollutant dispersion is attributed to model flaws and uncertainty contributed by closure coefficients. Previous studies have attempted to improve the performance of RANS modeling by ad hoc calibration of the turbulence model coefficients specifically for urban problems. However, these models failed to accurately reproduce the key features like the reattachment lengths. In addition, there was a lack of generalizability evaluations of the calibrated models. To optimize the accuracy and generalizability of the turbulence model, this study considered the effects of optimization objectives and model formulations. Six k-epsilon based turbulence models calibrated using the ensemble Kalman filtering (EnKF) approach were compared. A wind tunnel experiment consisting of key features of the airflow around a single-building model was conducted to provide the training data. The proposed optimization objective prioritizing the reproduction of the reattachment lengths in the calibrations enabled the calibrated models to capture the key features of the airflow with better accuracy. The six turbulence models before and after the calibrations were compared for a single-building test case and a building-array test case with respect to the reattachment lengths, velocity, turbulence kinetic energy, and airborne contaminant concentration. The calibrated models with different formulations exhibit distinctive generalizability. The calibrated Murakami-Mochida-Kondo k-epsilon model (MMK) exhibited strong potential for generalizability to single-building problems. However, the generalization from the single-building isolated roughness flow regime to the street canyon skimming flow regime is limited.

Automated summary

This study compared six k-epsilon based turbulence models calibrated using the ensemble Kalman filtering approach. The results showed that prioritizing the reproduction of reattachment lengths in the optimization objective improved the accuracy of the models in capturing key features. The calibrated models exhibited different levels of generalizability in different test cases, with the Murakami-Mochida-Kondo k-epsilon model showing strong potential for single-building problems.