On the suitability of second-order accurate finite-volume solvers for the simulation of atmospheric boundary layer flow

Abstract

Abstract. The present work analyzes the quality and reliability of an important class of general-purpose, second-order accurate finite-volume (FV) solvers for the large-eddy simulation of a neutrally stratified atmospheric boundary layer (ABL) flow. The analysis is carried out within the OpenFOAM® framework, which is based on a colocated grid arrangement. A series of open-channel flow simulations are carried out using a static Smagorinsky model for subgrid scale momentum fluxes in combination with an algebraic equilibrium wall-layer model. The sensitivity of the solution to variations in numerical parameters such as grid resolution (up to 1603 control volumes), numerical solvers, and interpolation schemes for the discretization of nonlinear terms is evaluated and results are contrasted against those from a well-established mixed pseudospectral–finite-difference code. Considered flow statistics include mean streamwise velocity, resolved Reynolds stresses, velocity skewness and kurtosis, velocity spectra, and two-point autocorrelations. A quadrant analysis along with the examination of the conditionally averaged flow field are performed to investigate the mechanisms responsible for momentum transfer in the flow. It is found that at the selected grid resolutions, the considered class of FV-based solvers yields a poorly correlated flow field and is not able to accurately capture the dominant mechanisms responsible for momentum transport in the ABL. Specifically, the predicted flow field lacks the well-known sweep and ejection pairs organized side by side along the cross-stream direction, which are representative of a streamwise roll mode. This is especially true when using linear interpolation schemes for the discretization of nonlinear terms. This shortcoming leads to a misprediction of flow statistics that are relevant for ABL flow applications and to an enhanced sensitivity of the solution to variations in grid resolution, thus calling for future research aimed at reducing the impact of modeling and discretization errors.

Development of a large-eddy simulation subgrid model based on artificial neural networks: a case study of turbulent channel flow

“…Park and Choi (2021) found, for instance, that using single-point inputs, rather than multi-point inputs, alleviated the observed a posteriori instability. Guan et al (2021) observed that convolutional neural networks achieved higher a priori accuracy than the multi-layer perceptron architecture selected in this study, and, interestingly, that the a posteriori stability depended on the number of training samples. Possibly, a corresponding further increase in the a priori accuracy…”

“…Several other studies (Guan et al, 2021;Park and Choi, 2021;Wang et al, 2018;Xie et al, 2019;Yang et al, 2019), in contrast, did report stable a posteriori results without requiring ad hoc adjustments, although in some cases only after using single-point rather than multi-point inputs (Park and Choi, 2021), or ensuring that sufficient training samples are presented (Guan et al, 2021).…”

“…This can possibly be overcome by applying known scalings and properties to the inputs and outputs (e.g. Ling et al, 2016b;Yang et al, 2019), extending the range of cases covered in the training data set, and/or retraining a previously optimized ANN on limited data from a new flow through transfer learning (Guan et al, 2021).…”

“…First of all, one way forward could be to adjust the training procedure such that it reflects better the a posteriori simulation. A potential elegant way to achieve this may be an online learning procedure similar to the ones proposed by Rasp (2020) and Guan et al (2021), where the ANN SGS model would be trained within the actual online LES simulation to reproduce the correct statistics and/or the correction terms inferred from a dynamically coupled DNS. Alternatively, errors expected to be introduced by an ANN SGS model in an a posteriori LES simulation could be added offline to the filtered flow fields u, v, w during training.…”

“…Some studies showed very promising results with this method, both in so-called a priori (offline) tests (where the predicted quantity is directly compared to the ones derived from DNS) and so-called a posteriori (online) tests (where the trained ANN is directly incorporated as a SGS model into a LES simulation). However, these studies mostly focused on 2-D/3-D (in)compressible isotropic turbulence (Beck et al, 2019;Guan et al, 2021;Maulik et al, 2019;Vollant et al, 2017;Wang et al, 2018;Xie et al, 2019;Zhou et al, 2019) and thus do not represent wall-bounded geoscientific flows. Furthermore, some of these studies (Beck et al, 2019;Maulik et al, 2019;Zhou et al, 2019) resorted to ad hoc adjustments (e.g.…”