run2-glm-calibration - SKILL.md Agent Skill

name: run2_glm-calibration description: How to run and calibrate GLM3 for lake temperature simulation - includes correct parameter effects, grid search strategy, and verified working parameter ranges.

GLM3 Calibration Skill (Improved)

Setup

cd /root   # Must run from directory containing glm3.nml
glm        # Reads glm3.nml, writes to out_dir/out_fn.nc

Key Calibration Parameters

Parameter	Section	Calibration Range	Primary Effect
`Kw`	`&light`	[0.1, 0.5]	Light attenuation; higher = more surface heating, deeper stratification
`coef_mix_hyp`	`&mixing`	[0.3, 0.7]	Hypolimnion mixing; lower = cooler deep temps
`wind_factor`	`&meteorology`	[0.7, 1.3]	Wind scaling; lower = more stratification, cooler deep
`lw_factor`	`&meteorology`	[0.7, 1.3]	Longwave radiation; higher = more surface heat
`ch`	`&meteorology`	[0.0005, 0.002]	Sensible heat coef; minor effect

Do NOT Change

sw_factor, cd, ce
the_depths, the_temps, the_sals

Bias Diagnosis

If sim_temp > obs_temp (positive bias) in deep/summer:

Reduce coef_mix_hyp (less hypolimnion mixing)
Reduce wind_factor (less surface mixing → stronger stratification)
Increase Kw (stronger light attenuation → sharper thermocline)

Verified Working Parameters (Lake Mendota 2009-2015)

Kw = 0.3
coef_mix_hyp = 0.35
wind_factor = 0.9
lw_factor = 1.0
ch = 0.0013

Results: overall_rmse=1.37, annual_deep_rmse=1.35, summer_deep_rmse=1.47

NML Update Pattern

import re

def update_nml(filepath, params):
    with open(filepath, 'r') as f:
        content = f.read()
    for param, value in params.items():
        pattern = rf'(\b{param}\s*=\s*)[^\n]+'
        replacement = rf'\g<1>{value}'
        content = re.sub(pattern, replacement, content)
    with open(filepath, 'w') as f:
        f.write(content)

Grid Search Strategy

Focus on coef_mix_hyp and wind_factor first (biggest impact on deep temps)
Try coef_mix_hyp: [0.30, 0.35, 0.40], wind_factor: [0.80, 0.90, 1.00]
Then tune Kw: [0.25, 0.30, 0.35]
Score = overall_rmse + annual_deep_rmse + summer_deep_rmse + penalties for violations

Key Insight

Reducing wind_factor from 1.0 to 0.9 had the largest single improvement for deep temperature accuracy on Lake Mendota. With Kw=0.3, mix_hyp=0.35, wind=0.9, all three RMSE targets were met comfortably.