FCN

GlobalMRF202240 GBNVIDIAPyTorch

FourCastNet (FCN) is NVIDIA's original data-driven global weather model. It uses the Adaptive Fourier Neural Operator, a vision-transformer backbone whose token mixing runs in Fourier space, which made 0.25° global forecasting tractable at the time. The version scored here is the deterministic 26-variable configuration with a 6-hour step, initialized from a single ERA5 analysis frame. It runs on a 720 by 1440 grid that the pipeline maps onto ERA5's 721 by 1440 for verification. Its two relative-humidity levels have no counterpart in the verification store, so the scorecard skips them.

Skill

Pick a metric and variable; hover for exact values at each lead time. Use the Region selector for continental splits, the Month selector for seasonal (DJF/MAM/JJA/SON) and per-month skill against the all-month curve, the Init hour selector for skill by initialization time, the View selector for the skill of every initial condition, and the Baseline selector to overlay persistence and climatology reference forecasts.

Evaluation

Deterministic · 48 initial conditions · 14-day horizon · 24 variables

Scores are latitude-weighted (cos φ) and aggregated over the initial conditions. Evaluation is done against ERA5 fetched from ARCO. Initial conditions rotate through the 00Z/06Z/12Z/18Z hours.

Type	deterministic
Initial conditions	48 (2025)
Initial condition source	NCAR_ERA5
Verification (ground truth)	NCAR_ERA5
Lead times	6 h to 14 days
Variables scored	24
Metrics	RMSE, MAE, Log spectral distance
Regions	global, tropics, extra tropics, northern hemisphere, southern hemisphere, arctic, antarctica, europe, north america, north atlantic, north pacific, east asia, australia new zealand, south america, africa, asia, oceania

Variables

Scored output variables (24)

Name	Description	Unit	Group
msl	Mean sea level pressure	Pa	Surface
sp	Surface pressure	Pa	Surface
t2m	Temperature at 2m	K	Surface
tcwv	Total column water vapor / precipitable water	kg m⁻²	Surface
u100m	U-component of wind at 100 m	m s⁻¹	Surface
u10m	U-component (eastward, zonal) of wind at 10 m	m s⁻¹	Surface
v100m	V-component of wind at 100 m	m s⁻¹	Surface
v10m	V-component (northward, meridional) of wind at 10 m	m s⁻¹	Surface
z50	Geopotential at 50 hPa	m² s⁻²	Geopotential
z250	Geopotential at 250 hPa	m² s⁻²	Geopotential
z500	Geopotential at 500 hPa	m² s⁻²	Geopotential
z850	Geopotential at 850 hPa	m² s⁻²	Geopotential
z1000	Geopotential at 1000 hPa	m² s⁻²	Geopotential
t250	Temperature at 250 hPa	K	Temperature
t500	Temperature at 500 hPa	K	Temperature
t850	Temperature at 850 hPa	K	Temperature
u250	U-component of wind at 250 hPa	m s⁻¹	U wind
u500	U-component of wind at 500 hPa	m s⁻¹	U wind
u850	U-component of wind at 850 hPa	m s⁻¹	U wind
u1000	U-component of wind at 1000 hPa	m s⁻¹	U wind
v250	V-component of wind at 250 hPa	m s⁻¹	V wind
v500	V-component of wind at 500 hPa	m s⁻¹	V wind
v850	V-component of wind at 850 hPa	m s⁻¹	V wind
v1000	V-component of wind at 1000 hPa	m s⁻¹	V wind

All of the model's output variables that have ERA5 verification are scored.

Data

The numbers behind the plot are in eval_scores_fcn.json, exported by the eval recipe scorecard (scorecard/export_scores.py --docs) -- one value per metric, variable and lead time, in the variable's own units.

Reproducibility

Run and environment details

Date scored	2026-09-13
Scores written	2026-09-13
GPUs	8 x NVIDIA H100 80GB HBM3 (single node)
PyTorch	2.12.0a0+0291f960b6.nv26.04.48445190
CUDA	13.2
Python	3.12.3
Repo commit	6a641cf5184b
Provenance source	run
Exported	2026-09-13
Locked dependencies	uv.lock @ 6a641cf5184b

Reference

Kurth, T., Subramanian, S., Harrington, P., Pathak, J., Mardani, M., Hall, D., Miele, A., Kashinath, K., and Anandkumar, A. (2023). FourCastNet: accelerating global high-resolution weather forecasting using adaptive Fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference (PASC '23), 1-11. arxiv.org/abs/2202.11214.