Turpy: A GPU-Native implicit LES Model for Meter-Scale Boundary Layer Turbulence Based on Energy-Conserving LMARS Dynamics

This study develops a novel implicit large-eddy simulation (ILES) model with strict energy-conserving properties, named Turpy, for high-resolution analysis of microscale turbulent structures in the atmospheric boundary layer. The model is implemented natively in Python and leverages the machine learning framework Pytorch to enable efficient GPU computation, achieving parallel scalability exceeding 90% at meter-scale spatial resolution. The numerics of Turpy is tailored towards the meter-scale turbulence simulation: first, the model adopts the energy-conserving form of the compressible Euler equations as the governing system, ensuring total energy conservation while naturally representing the conversion between internal and kinetic energy; second, the model employs a finite-volume discretization without introducing explicit scale filtering, and subgrid-scale effects are represented through the numerical dissipation generated by a Low Mach Number Approximate Riemann Solver (LMARS), eliminating the need for additional subgrid-scale turbulence parameterizations. Numerical experiments demonstrate that Turpy can reasonably reproduce the characteristic structures of wind and temperature fields in the boundary layer under different thermal stratification conditions. Furthermore, by incorporating a wind turbine model, Turpy accurately captures the spatial structure and evolution of wind turbine wakes, highlighting its strong capability and application potential in boundary-layer turbulence research and wind energy applications.