LES Intercomparison Study for Neutral Boundary Layers: SGS Model Coefficients
Last updated: May 2026
For case setup and physical parameters, see the Description notebook.
Vertical profiles of the SGS coefficients and scale-dependence parameters \(\beta_1\) (momentum) are compared across the four grid resolutions (\(64^3\), \(128^3\), \(256^3\), \(384^3\)) for a user-selected SGS model and precision.
Setup
The next cells load Python packages, locate the simulation outputs, and define the grid and averaging window used throughout the notebook.
[1]:
import os
import re
import glob
import numpy as np
import matplotlib.pyplot as plt
from pathlib import Path
Output directories
[2]:
from pathlib import Path
# Base directory (jaxalfa/)
def find_repo_root(start=None):
path = Path(start or ('__file__' in globals() and __file__) or Path.cwd()).resolve()
for candidate in (path, *path.parents):
if (candidate / 'examples').is_dir() and (candidate / 'docs').is_dir():
return candidate
raise FileNotFoundError('Could not locate jaxalfa repository root')
BaseDir = find_repo_root()
def read_config(run_dir):
cfg = {}
exec((run_dir / 'Config.py').read_text(), cfg)
return cfg
optSGS = 1 # LASDD-SM: 1, LASDD-WL: 2, LAD-SM: 3, LAD-WL: 4
optPrecision = 'DP' # 'DP' or 'SP'
sgs_names = {1: 'LASDD-SM', 2: 'LASDD-WL', 3: 'LAD-SM', 4: 'LAD-WL'}
run_styles = {
'64x64x64': {'color': 'red', 'linestyle': '-'},
'128x128x128': {'color': 'blue', 'linestyle': '-'},
'256x256x256': {'color': 'green', 'linestyle': '-'},
'384x384x384': {'color': 'black', 'linestyle': '-'},
}
_sgs = {1: 'LASDD_SM', 2: 'LASDD_WL', 3: 'LAD_SM', 4: 'LAD_WL'}
_run = lambda res: f'{res}_{_sgs[optSGS]}_{optPrecision}'
OutputDir1 = BaseDir / f'examples/NBL_A94/runs/{_run("64x64x64")}/output'
OutputDir2 = BaseDir / f'examples/NBL_A94/runs/{_run("128x128x128")}/output'
OutputDir3 = BaseDir / f'examples/NBL_A94/runs/{_run("256x256x256")}/output'
OutputDir4 = BaseDir / f'examples/NBL_A94/runs/{_run("384x384x384")}/output'
Case configuration
[3]:
cfg_1 = read_config(OutputDir1.parent)
cfg_2 = read_config(OutputDir2.parent)
cfg_3 = read_config(OutputDir3.parent)
cfg_4 = read_config(OutputDir4.parent)
nz_1 = int(cfg_1['nz'])
nz_2 = int(cfg_2['nz'])
nz_3 = int(cfg_3['nz'])
nz_4 = int(cfg_4['nz'])
l_z = float(cfg_1['l_z'])
z_damping = float(cfg_1.get('z_damping', np.nan))
OutputInterval_sec = float(cfg_1.get('OutputInterval_sec', 60.0))
# Averaging window — NBL quasi-steady state (last 10 h of an 83 h run)
T_start = 73 * 3600 # s
T_end = 83 * 3600 # s
Derived grid and averaging indices
[4]:
# Half levels — SGS coefficients live at UVP nodes
z_1 = np.array([(k + 0.5) * l_z / (nz_1 - 1) for k in range(nz_1)])
z_2 = np.array([(k + 0.5) * l_z / (nz_2 - 1) for k in range(nz_2)])
z_3 = np.array([(k + 0.5) * l_z / (nz_3 - 1) for k in range(nz_3)])
z_4 = np.array([(k + 0.5) * l_z / (nz_4 - 1) for k in range(nz_4)])
# File indices for the averaging window
T_start_index = int(T_start / OutputInterval_sec) - 1
T_end_index = int(T_end / OutputInterval_sec) - 1
print(f'Averaging window: file indices {T_start_index} – {T_end_index}')
Averaging window: file indices 4379 – 4979
SGS coefficient loader
[5]:
def LoadSGSAverage(stat_files, T_start_index, T_end_index, nz_expected):
"""Return time-averaged SGS coefficient profiles over the given window.
Returns
-------
Cs2_1, Cs2_2 : ndarray (nz)
Smagorinsky coefficient squared: (PlanarMean(Cs))^2 and PlanarMean(Cs^2).
Cs2PrRatio : ndarray (nz)
PlanarMean(Cs^2 / Pr_T) profile.
Beta1, Beta2 : ndarray (nz)
Scale-dependence parameters for momentum and scalar.
"""
if len(stat_files) == 0:
print(f'No statistics files available; plotting NaN placeholders for nz={nz_expected}.')
nan = np.full(nz_expected, np.nan)
return nan, nan.copy(), nan.copy(), nan.copy(), nan.copy()
sl = slice(T_start_index, min(T_end_index + 1, len(stat_files)))
if sl.start >= len(stat_files):
print(f'Averaging window starts after available files; plotting NaN placeholders for nz={nz_expected}.')
nan = np.full(nz_expected, np.nan)
return nan, nan.copy(), nan.copy(), nan.copy(), nan.copy()
Cs2_1_list = []; Cs2_2_list = []; Cs2PR_list = []
B1_list = []; B2_list = []
for f in stat_files[sl]:
with np.load(f) as d:
Cs2_1_list.append(d['Cs2_1'])
Cs2_2_list.append(d['Cs2_2'])
Cs2PR_list.append(d['Cs2PrRatio'])
B1_list.append(d['Beta1'])
B2_list.append(d['Beta2'])
return (
np.mean(Cs2_1_list, axis=0),
np.mean(Cs2_2_list, axis=0),
np.mean(Cs2PR_list, axis=0),
np.mean(B1_list, axis=0),
np.mean(B2_list, axis=0),
)
Available statistics files
[6]:
def get_stat_files(output_dir):
files = sorted(
glob.glob(str(output_dir / 'ALFA_Statistics_Iteration_*.npz')),
key=lambda x: int(re.search(r'Iteration_(\d+)', x).group(1))
)
return files
StatFiles1 = get_stat_files(OutputDir1)
StatFiles2 = get_stat_files(OutputDir2)
StatFiles3 = get_stat_files(OutputDir3)
StatFiles4 = get_stat_files(OutputDir4)
print(f'64^3 : {len(StatFiles1)} files')
print(f'128^3 : {len(StatFiles2)} files')
print(f'256^3 : {len(StatFiles3)} files')
print(f'384^3 : {len(StatFiles4)} files')
64^3 : 5000 files
128^3 : 5000 files
256^3 : 5000 files
384^3 : 0 files
Temporally averaged profiles
[7]:
(Cs2_1_avg_1, Cs2_2_avg_1, Cs2PR_avg_1, B1_avg_1, B2_avg_1) = \
LoadSGSAverage(StatFiles1, T_start_index, T_end_index, nz_1)
(Cs2_1_avg_2, Cs2_2_avg_2, Cs2PR_avg_2, B1_avg_2, B2_avg_2) = \
LoadSGSAverage(StatFiles2, T_start_index, T_end_index, nz_2)
(Cs2_1_avg_3, Cs2_2_avg_3, Cs2PR_avg_3, B1_avg_3, B2_avg_3) = \
LoadSGSAverage(StatFiles3, T_start_index, T_end_index, nz_3)
(Cs2_1_avg_4, Cs2_2_avg_4, Cs2PR_avg_4, B1_avg_4, B2_avg_4) = \
LoadSGSAverage(StatFiles4, T_start_index, T_end_index, nz_4)
# SGS coefficient: two averaging conventions
# Method 1: C = PlanarMean(C), then squared for storage → take sqrt
# Method 2: C = sqrt(PlanarMean(C^2))
Cs_m1_1 = np.sqrt(np.abs(Cs2_1_avg_1)); Cs_m2_1 = np.sqrt(np.abs(Cs2_2_avg_1))
Cs_m1_2 = np.sqrt(np.abs(Cs2_1_avg_2)); Cs_m2_2 = np.sqrt(np.abs(Cs2_2_avg_2))
Cs_m1_3 = np.sqrt(np.abs(Cs2_1_avg_3)); Cs_m2_3 = np.sqrt(np.abs(Cs2_2_avg_3))
Cs_m1_4 = np.sqrt(np.abs(Cs2_1_avg_4)); Cs_m2_4 = np.sqrt(np.abs(Cs2_2_avg_4))
# PrSGS = Cs2_2 / Cs2PrRatio (NaN where Cs2PrRatio ≈ 0)
_tol = 1e-10
PrSGS_1 = np.where(Cs2PR_avg_1 > _tol, Cs2_2_avg_1 / Cs2PR_avg_1, np.nan)
PrSGS_2 = np.where(Cs2PR_avg_2 > _tol, Cs2_2_avg_2 / Cs2PR_avg_2, np.nan)
PrSGS_3 = np.where(Cs2PR_avg_3 > _tol, Cs2_2_avg_3 / Cs2PR_avg_3, np.nan)
PrSGS_4 = np.where(Cs2PR_avg_4 > _tol, Cs2_2_avg_4 / Cs2PR_avg_4, np.nan)
print(f'Averaging over {T_end_index - T_start_index + 1} files '
f'({T_start/3600:.1f}–1{T_end/3600:.1f} h)')
No statistics files available; plotting NaN placeholders for nz=384.
Averaging over 601 files (73.0–183.0 h)
/var/folders/fm/088xfr2x1vs0yr7dmrdgjfrh0000gn/T/ipykernel_14282/2150077097.py:23: RuntimeWarning: divide by zero encountered in divide
PrSGS_1 = np.where(Cs2PR_avg_1 > _tol, Cs2_2_avg_1 / Cs2PR_avg_1, np.nan)
/var/folders/fm/088xfr2x1vs0yr7dmrdgjfrh0000gn/T/ipykernel_14282/2150077097.py:23: RuntimeWarning: invalid value encountered in divide
PrSGS_1 = np.where(Cs2PR_avg_1 > _tol, Cs2_2_avg_1 / Cs2PR_avg_1, np.nan)
/var/folders/fm/088xfr2x1vs0yr7dmrdgjfrh0000gn/T/ipykernel_14282/2150077097.py:24: RuntimeWarning: divide by zero encountered in divide
PrSGS_2 = np.where(Cs2PR_avg_2 > _tol, Cs2_2_avg_2 / Cs2PR_avg_2, np.nan)
/var/folders/fm/088xfr2x1vs0yr7dmrdgjfrh0000gn/T/ipykernel_14282/2150077097.py:24: RuntimeWarning: invalid value encountered in divide
PrSGS_2 = np.where(Cs2PR_avg_2 > _tol, Cs2_2_avg_2 / Cs2PR_avg_2, np.nan)
/var/folders/fm/088xfr2x1vs0yr7dmrdgjfrh0000gn/T/ipykernel_14282/2150077097.py:25: RuntimeWarning: divide by zero encountered in divide
PrSGS_3 = np.where(Cs2PR_avg_3 > _tol, Cs2_2_avg_3 / Cs2PR_avg_3, np.nan)
/var/folders/fm/088xfr2x1vs0yr7dmrdgjfrh0000gn/T/ipykernel_14282/2150077097.py:25: RuntimeWarning: invalid value encountered in divide
PrSGS_3 = np.where(Cs2PR_avg_3 > _tol, Cs2_2_avg_3 / Cs2PR_avg_3, np.nan)
[8]:
plt.rcParams.update({
"text.usetex": True,
"font.size": 14,
"axes.labelsize": 16,
"xtick.labelsize": 12,
"ytick.labelsize": 12
})
SGS Model Coefficient
For SM variants (LASDD-SM, LAD-SM) the coefficient is the Smagorinsky coefficient \(C_s\); for WL variants (LASDD-WL, LAD-WL) it is the Wong-Lilly SGS coefficient \(C\). Two planar-averaging conventions are shown:
Method 1 (left): \(C = \langle C \rangle_{xy}\) — planar mean of the pointwise field.
Method 2 (right): \(C = \sqrt{\langle C^2 \rangle_{xy}}\) — root of the planar mean of \(C^2\).
By Jensen’s inequality, Method 2 \(\geq\) Method 1.
[9]:
fig, axs = plt.subplots(1, 2, figsize=(10, 5), constrained_layout=True)
_coeff_sym = r'C_s' if optSGS in [1, 3] else r'C'
_coeff_name = 'Smagorinsky' if optSGS in [1, 3] else 'Wong-Lilly SGS'
for lbl, Cs_m1, Cs_m2, z in [
('64x64x64', Cs_m1_1, Cs_m2_1, z_1),
('128x128x128', Cs_m1_2, Cs_m2_2, z_2),
('256x256x256', Cs_m1_3, Cs_m2_3, z_3),
('384x384x384', Cs_m1_4, Cs_m2_4, z_4),
]:
style = run_styles[lbl]
axs[0].plot(Cs_m1, z, color=style['color'], linestyle=style['linestyle'], linewidth=2, label=lbl)
axs[1].plot(Cs_m2, z, color=style['color'], linestyle=style['linestyle'], linewidth=2, label=lbl)
axs[0].set_xlabel(rf"${_coeff_sym} = \langle {_coeff_sym} \rangle_{{xy}}$")
axs[0].set_ylabel(r"$z$ (m)")
axs[0].set_title("Method 1")
axs[1].set_xlabel(rf"${_coeff_sym} = \sqrt{{\langle {_coeff_sym}^2 \rangle_{{xy}}}}$")
axs[1].set_ylabel(r"$z$ (m)")
axs[1].set_title("Method 2")
for ax in axs:
ax.set_ylim(0, z_damping)
ax.grid()
ax.legend(frameon=False)
fig.suptitle(f"SGS Model Coefficient (73--83 h average): {_coeff_name} ({sgs_names[optSGS]} model) ({optPrecision})", fontsize=18)
plt.show()
Scale-dependence Parameter
The scale-dependence parameter \(\beta_1\) characterises how the dynamic momentum coefficient varies across filter scales. Values near unity indicate weak scale dependence; departures signal the model adapting to local turbulence structure.
[10]:
fig, ax = plt.subplots(figsize=(5, 6), constrained_layout=True)
for lbl, B1, z in [
('64x64x64', B1_avg_1, z_1),
('128x128x128', B1_avg_2, z_2),
('256x256x256', B1_avg_3, z_3),
('384x384x384', B1_avg_4, z_4),
]:
style = run_styles[lbl]
ax.plot(B1, z, color=style['color'], linestyle=style['linestyle'], linewidth=2, label=lbl)
ax.set_xlabel(r"$\beta_1$ (momentum)")
ax.set_ylabel(r"$z$ (m)")
ax.set_ylim(0, z_damping)
ax.grid()
ax.legend(frameon=False)
fig.suptitle(f"Momentum Scale-dependence Parameter (73--83 h average): {sgs_names[optSGS]} model ({optPrecision})", fontsize=18)
plt.show()
