A modern, unified framework for running and managing numerical models.

Designed for researchers who want to focus on science — not on the details of model execution.

wrfrun is a Python package that provides a general-purpose, reproducible, and extensible framework for running numerical models.

It automates the tedious parts of model execution — preparing input data, handling namelist configurations, organizing logs, and submitting jobs — so that you can spend your time on research, not on managing model runs.

pip install wrfrun

Check https://wrfrun.syize.cn.

wrfrun enforces a unified interface specification for all numerical models. Specifically, each model interface must inherit from a provided base class, ensuring a consistent structure and behavior across different models.

This design makes model execution intuitive — any supported model can be launched simply by calling a Python function or class method, while wrfrun automatically handles all background tasks such as data preparation and configuration file management.

from wrfrun.run import WRFRun
from wrfrun.model.wrf import geogrid, metgrid, real, ungrib, wrf

# wrfrun will prepare input data, generate namelist file, 
# save outputs and logs automatically.
with WRFRun("./config.toml", init_workspace=True) as wrf_run:
    geogrid()
    ungrib()
    metgrid()
    real()
    wrf()

Every simulation can be fully recorded and later reproduced from a single .replay file — ensuring total reproducibility.

from wrfrun.run import WRFRun
from wrfrun.model.wrf import geogrid, ungrib, metgrid, real, wrf


# 1. Record simulation with method `record_simulation`.
with WRFRun("./config.toml", init_workspace=True) as wrf_run:
    wrf_run.record_simulation(output_path="./outputs/example.replay")

    geogrid()
    ungrib()
    metgrid()
    real()
    wrf()

# 2. Replay the simulation in a different directory or on a different machine.
with WRFRun("./config.toml", init_workspace=True) as wrf_run:
    wrf_run.replay_simulation("./example.replay")

Manage all simulation settings in TOML files: A main config file, and model config files.

For more information about the configuration file, check config.

# main config file: config.toml
work_dir = "./.wrfrun"

input_data_path = ""
output_path = "./outputs"
log_path = "./logs"

server_host = "localhost"
server_port = 54321

core_num = 36

[job_scheduler]
job_scheduler = "pbs"

queue_name = ""
node_num = 1
env_settings = {}
python_interpreter = "/usr/bin/python3"     # or just "python3"

[model]
[model.wrf]
use = false
include = "./configs/wrf.toml"

wrfrun remains compatible with original namelist inputs, just set namelist file path in the model config.

# WRF model config file: wrf.toml
wps_path = '/path/to/your/WPS/folder'
wrf_path = '/path/to/your/WRF/folder'
wrfda_path = ''		# WRFDA is optional.
geog_data_path = '/path/to/your/geog/data'
user_wps_namelist = ''		# set your own namelist file here
user_real_namelist = ''		# set your own namelist file here
user_wrf_namelist = ''		# set your own namelist file here
user_wrfda_namelist = ''	# set your own namelist file here
restart_mode = false
debug_level = 100

[time]
start_date = 2021-03-24T12:00:00Z   # or [2021-03-24T12:00:00Z, 2021-03-24T12:00:00Z]
end_date = 2021-03-26T00:00:00Z     # or [2021-03-26T00:00:00Z, 2021-03-24T12:00:00Z]
input_data_interval = 10800
output_data_interval = 180
time_step = 120
parent_time_step_ratio = [1, 3, 4]
restart_interval = -1


[domain]
domain_num = 3
parent_grid_ratio = [1, 3, 9]
i_parent_start = [1, 17, 72]
j_parent_start = [1, 17, 36]
e_we = [120, 250, 1198]
e_sn = [120, 220, 1297]
dx = 9000
dy = 9000
map_proj = 'lambert'
truelat1 = 34.0
truelat2 = 40.0
ref_lat = 37.0
ref_lon = 120.5
stand_lon = 120.5


[scheme]
long_wave_scheme = { name = "rrtm", option = {} }
short_wave_scheme = { name = "rrtmg", option = {} }
cumulus_scheme = { name = "kf", option = {} }
pbl_scheme = { name = "ysu", option = { ysu_topdown_pblmix = 1} }
land_surface_scheme = { name = "noah", option = {} }
surface_layer_scheme = { name = "mm5", option = {} }

Automatically submit jobs to supported schedulers:

• PBS
• Slurm
• LSF

wrfrun takes care of resource requests and queue management automatically.

from wrfrun.run import WRFRun
from wrfrun.model.wrf import geogrid, metgrid, real, ungrib, wrf

# just set submit_job=True
with WRFRun("./config.toml", init_workspace=True, submit_job=True) as wrf_run:
    geogrid()
    ungrib()
    metgrid()
    real()
    wrf()

wrfrun can parse model log files and start a lightweight socket server to report simulation progress.

from wrfrun.run import WRFRun
from wrfrun.model.wrf import geogrid, metgrid, real, ungrib, wrf

# just set start_server=True
with WRFRun("./config.toml", init_workspace=True, start_server=True) as wrf_run:
    geogrid()
    ungrib()
    metgrid()
    real()
    wrf()

• Automated ERA5 data download (requires cdsapi authentication)
• Real-time progress reporting via socket interface
• Numerical model support:
  • Partial WRF support:
    • Full support for WPS
    • Wrapped execution for real and wrf
  • PALM model support.
• Job submission on PBS, Slurm, and LSF
• record / replay reproducibility for all compliant interfaces

• Full WRF model integration.
• Broaden model support.
• Enhanced progress visualization dashboard.

This project is currently for personal and research use. If you have ideas or feature requests, feel free to open an issue.