---
license: cc-by-4.0
metrics:
- mse
- csi
pipeline_tag: image-to-image
language:
- en
library_name: pytorch
tags:
- weather
- weather-forecasting
- climate
- machine-learning-weather-prediction
- fuxi
- transformer
---
# FuXi-2.1
[](https://creativecommons.org/licenses/by/4.0/)
[]()
[]()
**FuXi-2.1** is a global, deterministic machine-learning weather forecasting model
developed by **Fudan University** & **[SAIS](https://www.sais.com.cn/)**.
It produces global forecasts at **0.25°** resolution, on **6-hourly** steps, out to
**10 days**.
FuXi-2.1 targets the defining failure mode of data-driven weather prediction: forecasts
that blur into a smooth spatial average as lead time grows, erasing the small-scale
structure that matters most for extremes. FuXi-2.1 produces markedly **sharper** fields
whose spatial power spectra track observations across the full wavenumber range, while
keeping deterministic skill (RMSE) **comparable to** FuXi-1.0 — and substantially
improving extreme-event detection for heavy precipitation and strong wind.
This model is released as part of the **FuXi Single** collection.
## Table of contents
- [What's new in 2.1](#whats-new-in-21)
- [Quickstart](#quickstart)
- [Model overview](#model-overview)
- [Data details](#data-details)
- [Evaluation](#evaluation)
- [Known limitations](#known-limitations)
- [Citation](#citation)
---
## What's new in 2.1
Relative to FuXi-1.0, FuXi-2.1 introduces:
- **A flat Transformer backbone** replacing FuXi-1.0's U-Transformer (ResNet
downsample → Swin Transformer → upsample). FuXi-2.1 drops the U-shaped
up/down-sampling in favour of a single full-resolution Transformer trunk.
- **Rotary position embeddings (RoPE)** inside the Swin windowed attention, replacing
the learned relative-position bias used in FuXi-1.0.
- **adaLN time conditioning** that injects time-period information — forecast lead step,
time-of-day and day-of-year phase — into every block, inspired by diffusion
transformers in image generation.
- **A variable-aware multi-head decoder** that gives pressure-level, surface and derived
variables their own specialised output heads.
The combined effect is sharper, spectrally faithful forecasts with no penalty on
mean-error skill.
---
## Quickstart
This repository ships the exported model (`fuxi-2.1.pt2`), normalization statistics
(`mean.nc`, `std.nc`), a sample pre-normalized input (`input.nc`), and minimal inference
code.
```bash
# 1. Install dependencies
pip install -r requirements.txt
# 2. End-to-end demo (inference + plots)
bash run.sh --model_dir . --input input.nc --steps 5
# Or run inference directly (40 steps = 10-day forecast):
python inference.py \
--model_dir . \
--input input.nc \
--output_dir ./output \
--steps 40 \
--forecast_time 2024092900
# 3. Plot selected channels
python plot.py --output_dir ./output --channels t2m z500 tp --discrete
```
**Input** — a NetCDF with a variable `input` of shape `(time=2, channel=85, lat=721,
lon=1440)`, z-score normalized, coordinates `lat` 90→−90 and `lon` 0→359.75. The provided
`input.nc` is a sample for 2024-09-29 00Z.
**Output** — each step saved as `{output_dir}/{step:03d}.nc`, shape `(channel=85, lat=721,
lon=1440)` in physical units (denormalized), with a `valid_time` attribute. Steps are
1-based: `001.nc` = +6 h, … `040.nc` = +240 h (10 days).
> **GPU:** the device is baked into the exported graph; load on CUDA. ~8 GB GPU memory is
> enough (model ~4 GB + recurrent state ~1.4 GB + working memory). Tested on A100, V100,
> RTX 3090/4090. See `variables.py` for the full ordered channel list.
---
## Model overview
### Model description
FuXi-2.1 is a single Transformer. The global atmospheric state is split into patches and
embedded into tokens, processed by a stack of windowed-attention blocks, and read out by
a variable-aware multi-head decoder. The model is **deterministic** — one forward pass
per step, with no adversarial or diffusion sampling at inference — and is rolled out
**autoregressively** at 6-hourly steps.
- **Developed by:** Fudan University & [SAIS](https://www.sais.com.cn/)
- **Model type:** Transformer (patch-embed → Swin attention with RoPE + adaLN → multi-head decoder)
- **Forecast type:** Global, deterministic, autoregressive
- **License:** CC BY 4.0
- **Predecessor:** [FuXi-1.0](https://github.com/tpys/FuXi)
### Architecture details
| Component | Specification |
|---|---|
| Backbone | Single Transformer trunk (no U-Net up/down-sampling) |
| Attention | Swin windowed attention |
| Position encoding | Rotary (RoPE, 1-D) |
| Normalisation / conditioning | adaLN, conditioned on lead step, time-of-day, day-of-year |
| Feed-forward | SwiGLU |
| Decoder | Variable-aware multi-head (pressure / surface / derived) |
| Input frames | 2 (states at t−6h and t₀) |
| Output | State at t+6h, rolled out autoregressively |
### Model resolution
| Model | Horizontal resolution | Vertical resolution [pressure levels] (hPa) |
|:---|:---:|:---|
| FuXi-2.1 | 0.25° (721×1440) | 13: 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 |
---
## Data details
### Training data
FuXi-2.1 is trained and evaluated on **ERA5** reanalysis at 0.25° resolution, 6-hourly.
- **Training period:** 2002–2023
- **Test period:** 2024 (held out)
### Data parameters
FuXi-2.1 operates on **85 channels** per time step: **65 pressure-level** channels
(5 variables × 13 levels) and **20 surface** channels, plus static forcings supplied as
constant inputs. Most channels are **prognostic** — the same channels are input and
output and fed back during roll-out. Radiation fluxes and total precipitation are
**diagnostic** outputs produced through a dedicated decoder head (they are predicted but
not fed back as inputs).
Channel order (exact): the 65 pressure-level channels first (z, then t, u, v, q, each
over the 13 levels 50→1000 hPa), followed by the 20 surface channels:
`msl, t2m, d2m, sst, ws10m, ws100m, u10m, v10m, u100m, v100m, lcc, mcc, hcc, tcc, ssr,
ssrd, fdir, ttr, tcw, tp`.
#### Pressure-level parameters (13 levels: 50–1000 hPa)
| Short name | Name | Units | Input/Output |
|:---:|:---|:---:|:---:|
| z | Geopotential | m²·s⁻² | Both (prognostic) |
| t | Temperature | K | Both (prognostic) |
| u | Eastward wind | m·s⁻¹ | Both (prognostic) |
| v | Northward wind | m·s⁻¹ | Both (prognostic) |
| q | Specific humidity | kg·kg⁻¹ | Both (prognostic) |
#### Surface parameters (20)
| Short name | Name | Units | Input/Output |
|:---:|:---|:---:|:---:|
| msl | Mean sea-level pressure | Pa | Both (prognostic) |
| t2m | 2 m temperature | K | Both (prognostic) |
| d2m | 2 m dewpoint temperature | K | Both (prognostic) |
| sst | Sea-surface temperature | K | Both (prognostic) |
| ws10m | 10 m wind speed | m·s⁻¹ | Both (prognostic) |
| ws100m | 100 m wind speed | m·s⁻¹ | Both (prognostic) |
| u10m | 10 m eastward wind | m·s⁻¹ | Both (prognostic) |
| v10m | 10 m northward wind | m·s⁻¹ | Both (prognostic) |
| u100m | 100 m eastward wind | m·s⁻¹ | Both (prognostic) |
| v100m | 100 m northward wind | m·s⁻¹ | Both (prognostic) |
| lcc | Low cloud cover | 0–1 | Both (prognostic) |
| mcc | Medium cloud cover | 0–1 | Both (prognostic) |
| hcc | High cloud cover | 0–1 | Both (prognostic) |
| tcc | Total cloud cover | 0–1 | Both (prognostic) |
| tcw | Total column water | kg·m⁻² | Both (prognostic) |
| ssr | Surface net solar radiation | J·m⁻² | Output (diagnostic) |
| ssrd | Surface solar radiation downwards | J·m⁻² | Output (diagnostic) |
| fdir | Total-sky direct solar radiation at surface | J·m⁻² | Output (diagnostic) |
| ttr | Top net thermal radiation | J·m⁻² | Output (diagnostic) |
| tp | Total precipitation | mm | Output (diagnostic) |
| Field | Level type | Input/Output |
|---|---|---|
| Land-sea mask, orography/geopotential, latitude/longitude encodings, time-of-day / day-of-year | Surface / static | Input (forcings) |
---
## Evaluation
We compare FuXi-2.1 against FuXi-1.0 under an identical protocol: forecasts initialised
from ERA5 and rolled out to 240 h in 6-hour steps. CSI is computed over **land only,
globally**. These numbers come from a **limited set of sample cases**, not a full-year
evaluation — they are indicative, and broader scorecards will follow.
**Headline:** RMSE stays comparable to FuXi-1.0 across variables, while structural and
extreme-event scores improve substantially.
### Precipitation — Critical Success Index (CSI)
| Threshold | FuXi-1.0 | FuXi-2.1 | Δ |
|:---|---:|---:|---:|
| ≥ 5 mm | 0.265 | 0.284 | +7.3% |
| ≥ 20 mm | 0.131 | 0.146 | +11.4% |
| ≥ 50 mm | 0.074 | 0.084 | +13.4% |
| ≥ 100 mm | 0.014 | 0.024 | **+68.3%** |
### 10 m wind speed — Critical Success Index (CSI)
| Threshold | FuXi-1.0 | FuXi-2.1 | Δ |
|:---|---:|---:|---:|
| ≥ 10.8 m·s⁻¹ | 0.544 | 0.571 | +4.8% |
| ≥ 24.5 m·s⁻¹ | 0.165 | 0.198 | +20.3% |
| ≥ 28.5 m·s⁻¹ | 0.000 | 0.044 | newly resolved |
The relative gain grows with event intensity, peaking at the extreme tail. At the
28.5 m·s⁻¹ wind threshold FuXi-1.0 scores zero — it never predicts such winds — whereas
FuXi-2.1 attains a non-zero CSI. Spatial power spectra of FuXi-2.1 track the observed
spectra across the full wavenumber range, in contrast to FuXi-1.0's high-wavenumber
energy deficit.
---
## Known limitations
- FuXi-2.1 is a deterministic model; it does not provide a calibrated ensemble spread.
- The CSI numbers reported here are computed on land only, over a limited set of sample
cases rather than a full-year evaluation; treat them as indicative. Comprehensive
global scorecards will be added.
- As with all ERA5-trained models, skill depends on the quality and resolution of the
initial conditions.
---
## Citation
If you use FuXi-2.1, please cite the FuXi series:
```bibtex
@article{chen2023fuxi,
title = {FuXi: a cascade machine learning forecasting system for 15-day global weather forecast},
author = {Chen, Lei and Zhong, Xiaohui and Zhang, Feng and Cheng, Yuan and Xu, Yimin and Qi, Yan and Li, Hao},
journal = {npj Climate and Atmospheric Science},
year = {2023},
volume = {6},
number = {1},
pages = {190}
}
```
---
**Code:** FuXi-1.0 — https://github.com/tpys/FuXi
*© 2026 Fudan University & [SAIS](https://www.sais.com.cn/) · FuXi Weather.*
