ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction
ChaosBench is a benchmark project to improve and extend the predictability range of deep weather emulators to the subseasonal-to-seasonal (S2S) range. Predictability at this scale is more challenging due to its: (1) double sensitivities to intial condition (in weather-scale) and boundary condition (in climate-scale), (2) butterfly effect, and our (3) inherent lack of understanding of physical processes operating at this scale. Thus, given the high socioeconomic stakes for accurate, reliable, and stable S2S forecasts (e.g., for disaster/extremes preparedness), this benchmark is timely for DL-accelerated solutions.
✨ Features
1️⃣ Diverse Observations. Spanning over 45 years (1979 - 2023), we include ERA5/LRA5/ORAS5 reanalysis for a fully-coupled Earth system emulation (atmosphere-terrestrial-sea-ice)
2️⃣ Diverse Baselines. Wide selection of physics-based forecasts from leading national agencies in Europe, the UK, America, and Asia
3️⃣ Differentiable Physics Metrics. Introduces two differentiable physics-based metrics to minimize the decay of power spectra at long forecasting horizon (blurriness)
4️⃣ Large-Scale Benchmarking. Systematic evaluation (deterministic, probabilistic, physics-based) for state-of-the-art ML-based weather emulators like ViT/ClimaX, PanguWeather, GraphCast, and FourcastNetV2
🏁 Getting Started
NOTE: Only need the dataset? Jump directly to Step 2. If you find any problems, feel free to contact us or raise a GitHub issue.
Step 0: Clone the ChaosBench Github repository
Step 1: Install package dependencies
$ cd ChaosBench
$ pip install -r requirements.txt
Step 2: Initialize the data space by running
$ cd data/
$ wget https://huggingface.co/datasets/LEAP/ChaosBench/resolve/main/process.sh
$ chmod +x process.sh
Step 3: Download the data
# Required for inputs and climatology (e.g., normalization)
$ ./process.sh era5
$ ./process.sh lra5
$ ./process.sh oras5
$ ./process.sh climatology
# Optional: control (deterministic) forecasts
$ ./process.sh ukmo
$ ./process.sh ncep
$ ./process.sh cma
$ ./process.sh ecmwf
# Optional: perturbed (ensemble) forecasts
$ ./process.sh ukmo_ensemble
$ ./process.sh ncep_ensemble
$ ./process.sh cma_ensemble
$ ./process.sh ecmwf_ensemble
🔍 Dataset Overview
All data has daily and 1.5-degree resolution.
ERA5 Reanalysis for Surface-Atmosphere (1979-2023). The following table indicates the 48 variables (channels) that are available for Physics-based models. Note that the Input ERA5 observations contains ALL fields, including the unchecked boxes:
Parameters/Levels (hPa) 1000 925 850 700 500 300 200 100 50 10 Geopotential height, z ($gpm$) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Specific humidity, q ($kg kg^{-1}$) ✓ ✓ ✓ ✓ ✓ ✓ ✓ Temperature, t ($K$) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ U component of wind, u ($ms^{-1}$) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ V component of wind, v ($ms^{-1}$) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Vertical velocity, w ($Pas^{-1}$) ✓ LRA5 Reanalysis for Terrestrial (1979-2023)
| Acronyms | Long Name | Units |
|---|---|---|
| asn | snow albedo | (0 - 1) |
| d2m | 2-meter dewpoint temperature | K |
| e | total evaporation | m of water equivalent |
| es | snow evaporation | m of water equivalent |
| evabs | evaporation from bare soil | m of water equivalent |
| evaow | evaporation from open water | m of water equivalent |
| evatc | evaporation from top of canopy | m of water equivalent |
| evavt | evaporation from vegetation transpiration | m of water equivalent |
| fal | forecaste albedo | (0 - 1) |
| lai_hv | leaf area index, high vegetation | $m^2 m^{-2}$ |
| lai_lv | leaf area index, low vegetation | $m^2 m^{-2}$ |
| pev | potential evaporation | m |
| ro | runoff | m |
| rsn | snow density | $kg m^{-3}$ |
| sd | snow depth | m of water equivalent |
| sde | snow depth water equivalent | m |
| sf | snowfall | m of water equivalent |
| skt | skin temperature | K |
| slhf | surface latent heat flux | $J m^{-2}$ |
| smlt | snowmelt | m of water equivalent |
| snowc | snowcover | % |
| sp | surface pressure | Pa |
| src | skin reservoir content | m of water equivalent |
| sro | surface runoff | m |
| sshf | surface sensible heat flux | $J m^{-2}$ |
| ssr | net solar radiation | $J m^{-2}$ |
| ssrd | download solar radiation | $J m^{-2}$ |
| ssro | sub-surface runoff | m |
| stl1 | soil temperature level 1 | K |
| stl2 | soil temperature level 2 | K |
| stl3 | soil temperature level 3 | K |
| stl4 | soil temperature level 4 | K |
| str | net thermal radiation | $J m^{-2}$ |
| strd | downward thermal radiation | $J m^{-2}$ |
| swvl1 | volumetric soil water layer 1 | $m^3 m^{-3}$ |
| swvl2 | volumetric soil water layer 2 | $m^3 m^{-3}$ |
| swvl3 | volumetric soil water layer 3 | $m^3 m^{-3}$ |
| swvl4 | volumetric soil water layer 4 | $m^3 m^{-3}$ |
| t2m | 2-meter temperature | K |
| tp | total precipitation | m |
| tsn | temperature of snow layer | K |
| u10 | 10-meter u-wind | $ms^{-1}$ |
| v10 | 10-meter v-wind | $ms^{-1}$ |
- ORAS Reanalysis for Sea-Ice (1979-2023)
| Acronyms | Long Name | Units |
|---|---|---|
| iicethic | sea ice thickness | m |
| iicevelu | sea ice zonal velocity | $ms^{-1}$ |
| iicevelv | sea ice meridional velocity | $ms^{-1}$ |
| ileadfra | sea ice concentration | (0-1) |
| so14chgt | depth of 14$^\circ$ isotherm | m |
| so17chgt | depth of 17$^\circ$ isotherm | m |
| so20chgt | depth of 20$^\circ$ isotherm | m |
| so26chgt | depth of 26$^\circ$ isotherm | m |
| so28chgt | depth of 28$^\circ$ isotherm | m |
| sohefldo | net downward heat flux | $W m^{-2}$ |
| sohtc300 | heat content at upper 300m | $J m^{-2}$ |
| sohtc700 | heat content at upper 700m | $J m^{-2}$ |
| sohtcbtm | heat content for total water column | $J m^{-2}$ |
| sometauy | meridonial wind stress | $N m^{-2}$ |
| somxl010 | mixed layer depth 0.01 | m |
| somxl030 | mixed layer depth 0.03 | m |
| sosaline | salinity | Practical Salinity Units (PSU) |
| sossheig | sea surface height | m |
| sosstsst | sea surface temperature | $^\circ C$ |
| sowaflup | net upward water flux | $kg/m^2/s$ |
| sozotaux | zonal wind stress | $N m^{-2}$ |
💡 Baseline Models
In addition to climatology and persistence, we evaluate the following:
- Physics-based models (including control/perturbed forecasts):
- UKMO: UK Meteorological Office
- NCEP: National Centers for Environmental Prediction
- CMA: China Meteorological Administration
- ECMWF: European Centre for Medium-Range Weather Forecasts
- Data-driven models:
- Lagged-Autoencoder
- Fourier Neural Operator (FNO)
- ResNet
- UNet
- ViT/ClimaX
- PanguWeather
- GraphCast
- Fourcastnetv2
🏅 Evaluation Metrics
We divide our metrics into 3 classes: (1) Deterministic-based, which cover evaluation used in conventional deterministic forecasting tasks, (2) Physics-based, which are aimed to construct a more physically-faithful and explainable data-driven forecast, and (3) Probabilistic-based, which account for the skillfulness of ensemble forecasts.
Deterministic-based:
- RMSE
- Bias
- Anomaly Correlation Coefficient (ACC)
- Multiscale Structural Similarity Index (MS-SSIM)
Physics-based:
- Spectral Divergence (SpecDiv)
- Spectral Residual (SpecRes)
Probabilistic-based:
- RMSE Ensemble
- Bias Ensemble
- ACC Ensemble
- MS-SSIM Ensemble
- SpecDiv Ensemble
- SpecRes Ensemble
- Continuous Ranked Probability Score (CRPS)
- Continuous Ranked Probability Skill Score (CRPSS)
- Spread
- Spread/Skill Ratio
🪜 Leaderboard
You can access the full score and checkpoints in logs/<MODEL_NAME> within the following subdirectory:
- Scores:
eval/<METRIC>.csv - Model checkpoints:
lightning_logs/
- Downloads last month
- 2