Reconstruction of unstable atmospheric surface layer streamwise turbulence based on multi-layer perceptron neural network architecture

Abstract

The accurate simulation of sand-laden turbulence under different stratification stabilities remains a critical challenge in turbulence research. This study presents an innovative approach to reconstructing streamwise turbulence in an unstable atmospheric surface layer (ASL) using a multi-layer perceptron (MLP) neural network architecture. Leveraging high-resolution measurements of three-dimensional wind velocity and temperature from multiple observational sites, the study develops prediction models for streamwise wind velocity at varying heights in the unstable ASL. The predictive model integrates large-scale motions (LSMs) generated by the MLP, small-scale motions (SSMs) derived from the Kaimal spectrum, and mean wind velocity, providing a comprehensive representation of turbulence. The impact of sand content and stratification stability on model performance is analyzed, with discussion highlighting the model's strengths and limitations under weak instability conditions. Validation is conducted through cross-site comparison, statistical analysis, and power spectrum assessment, demonstrating the model's ability to capture the temporal and spectral characteristics of wind velocity in sand-laden, unstable ASL conditions. The study also reveals that, under weak instability, shear forces dominate the formation of coherent structures, while buoyancy effects enhance vertical mixing as instability increases. Compared to existing models, the proposed prediction model is applicable over a broader range of conditions, offering a valuable data source for the study of atmospheric sand-laden turbulence and serving as a reference for practical sand control projects.