Surface‐Constrained Reconstruction of Spatiotemporal Turbulence Dissipation in the Convective Boundary Layer Using Physics‐Guided Machine Learning

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Turbulence dissipation rate (EDR) governs the conversion of turbulent kinetic energy to heat and thereby controls vertical mixing, transport, and energy exchange in the convective boundary layer (CBL). The spatiotemporal structure of EDR remains poorly constrained, as observations are limited to point measurements or sparse vertical profiles that cannot resolve spatiotemporal evolution, hindering evaluation of planetary boundary layer (PBL) parameterizations that rely on simplified turbulence representations. Here we develop a surface‐constrained framework that reconstructs CBL EDR structure from routine surface observations by integrating mixed‐layer physical constraints with a machine‐learning mapping trained on Doppler‐lidar–retrieved EDR at the ARM SGP site. The framework explains ∼70% of the observed EDR variance and reproduces its spatiotemporal structure. Model skill decreases systematically with height ( R 2 0.7 to 0.5 from surface to CBL top), consistent with the transition from surface‐driven to entrainment‐zone dynamics; a height‐resolved feature‐attribution analysis confirms that this decay tracks the diminishing importance of surface‐forcing predictors with height ( r = 0.78). Seasonal analysis shows highest predictability in autumn ( R 2 ≈ 0.75) and reduced skill in winter, reflecting regime‐dependent turbulence production. Feature attribution shows that surface heat flux and near‐surface turbulence scales dominate EDR variability, recovering classical similarity scaling through data‐driven inference. Cross‐site evaluation confirms generalization across nearby locations ( R 2 ≈ 0.67), indicating that surface forcing is a robust first‐order constraint on CBL turbulence structure. Remaining midday convective biases highlight entrainment and shear processes not captured by surface inputs. These results supply a data‐driven benchmark for evaluating and improving PBL parameterizations.
Funding Information
  • Office of Naval Research (N00014‐24‐1‐2554)
  • Biological and Environmental Research (DE‐AC05‐76RL01830)

This publication has 34 references indexed in Scilit: