Update README.md
Browse files
README.md
CHANGED
@@ -3,51 +3,63 @@ license: gpl-3.0
|
|
3 |
viewer: false
|
4 |
---
|
5 |
# ChaosBench
|
6 |
-
|
7 |
-
It is framed as a high-dimensional video regression task that consists of 45-year, 60-channel observations
|
8 |
-
for validating physics-based and data-driven models, and training the latter.
|
9 |
-
Physics-based forecasts are generated from 4 national weather agencies with 44-day lead-time and serve as baselines to data-driven forecasts.
|
10 |
-
Our benchmark is one of the first to incorporate physics-based metrics to ensure physically-consistent and explainable models.
|
11 |
-
We establish two tasks: full and sparse dynamics prediction.
|
12 |
|
13 |
-
|
14 |
|
15 |
π: [https://arxiv.org/abs/2402.00712](https://arxiv.org/abs/2402.00712)
|
16 |
|
17 |
-
##
|
18 |
-
**Step 1**: Clone the [ChaosBench](https://github.com/leap-stc/ChaosBench) Github repository
|
19 |
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
```
|
22 |
-
cd ChaosBench
|
23 |
-
pip install -r requirements.txt
|
24 |
```
|
25 |
|
26 |
-
**Step
|
27 |
```
|
28 |
-
cd data/
|
29 |
-
wget https://huggingface.co/datasets/LEAP/ChaosBench/resolve/main/process.sh
|
30 |
-
chmod +x process.sh
|
31 |
```
|
32 |
-
**Step
|
33 |
-
|
34 |
```
|
35 |
-
#
|
36 |
-
|
37 |
-
./process.sh
|
38 |
-
./process.sh
|
39 |
-
./process.sh
|
40 |
-
|
41 |
-
|
42 |
-
./process.sh
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
```
|
44 |
|
45 |
-
## Dataset Overview
|
|
|
|
|
|
|
46 |
|
47 |
-
- __Input:__ ERA5 Reanalysis (1979-2023)
|
48 |
-
|
49 |
-
- __Target:__ 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:
|
50 |
-
|
51 |
Parameters/Levels (hPa) | 1000 | 925 | 850 | 700 | 500 | 300 | 200 | 100 | 50 | 10
|
52 |
:---------------------- | :----| :---| :---| :---| :---| :---| :---| :---| :--| :-|
|
53 |
Geopotential height, z ($gpm$) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
|
@@ -57,8 +69,84 @@ chmod +x process.sh
|
|
57 |
V component of wind, v ($ms^{-1}$) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
|
58 |
Vertical velocity, w ($Pas^{-1}$) | | | | | ✓ | | | | | |
|
59 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
- __Baselines:__
|
61 |
-
- Physics-based models:
|
62 |
- [x] UKMO: UK Meteorological Office
|
63 |
- [x] NCEP: National Centers for Environmental Prediction
|
64 |
- [x] CMA: China Meteorological Administration
|
@@ -70,23 +158,36 @@ chmod +x process.sh
|
|
70 |
- [x] UNet
|
71 |
- [x] ViT/ClimaX
|
72 |
- [x] PanguWeather
|
73 |
-
- [x] Fourcastnetv2
|
74 |
- [x] GraphCast
|
|
|
75 |
|
76 |
-
## Evaluation Metrics
|
77 |
-
We divide our metrics into
|
78 |
|
79 |
-
|
|
|
80 |
- [x] RMSE
|
81 |
- [x] Bias
|
82 |
- [x] Anomaly Correlation Coefficient (ACC)
|
83 |
- [x] Multiscale Structural Similarity Index (MS-SSIM)
|
84 |
-
|
85 |
- [x] Spectral Divergence (SpecDiv)
|
86 |
- [x] Spectral Residual (SpecRes)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
|
89 |
-
## Leaderboard
|
90 |
You can access the full score and checkpoints in `logs/<MODEL_NAME>` within the following subdirectory:
|
91 |
- Scores: `eval/<METRIC>.csv`
|
92 |
- Model checkpoints: `lightning_logs/`
|
|
|
3 |
viewer: false
|
4 |
---
|
5 |
# ChaosBench
|
6 |
+
ChaosBench is a benchmark project to improve long-term forecasting of chaotic systems, in particular subseasonal-to-seasonal (S2S) climate, using ML approaches.
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
+
π: [https://leap-stc.github.io/ChaosBench/](https://leap-stc.github.io/ChaosBench/)
|
9 |
|
10 |
π: [https://arxiv.org/abs/2402.00712](https://arxiv.org/abs/2402.00712)
|
11 |
|
12 |
+
## β¨ Features
|
|
|
13 |
|
14 |
+
1οΈβ£ __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)
|
15 |
+
|
16 |
+
2οΈβ£ __Diverse Baselines__. Wide selection of physics-based forecasts from leading national agencies in Europe, the UK, America, and Asia
|
17 |
+
|
18 |
+
3οΈβ£ __Differentiable Physics Metrics__. Introduces two differentiable physics-based metrics to minimize the decay of power spectra at long forecasting horizon (blurriness)
|
19 |
+
|
20 |
+
4οΈβ£ __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
|
21 |
+
|
22 |
+
## π Getting Started
|
23 |
+
**Step 0**: Clone the [ChaosBench](https://github.com/leap-stc/ChaosBench) Github repository
|
24 |
+
|
25 |
+
**Step 1**: Install package dependencies
|
26 |
```
|
27 |
+
$ cd ChaosBench
|
28 |
+
$ pip install -r requirements.txt
|
29 |
```
|
30 |
|
31 |
+
**Step 2**: Initialize the data space by running
|
32 |
```
|
33 |
+
$ cd data/
|
34 |
+
$ wget https://huggingface.co/datasets/LEAP/ChaosBench/resolve/main/process.sh
|
35 |
+
$ chmod +x process.sh
|
36 |
```
|
37 |
+
**Step 3**: Download the data
|
|
|
38 |
```
|
39 |
+
# Required for inputs and climatology (e.g., normalization)
|
40 |
+
$ ./process.sh era5
|
41 |
+
$ ./process.sh lra5
|
42 |
+
$ ./process.sh oras5
|
43 |
+
$ ./process.sh climatology
|
44 |
+
|
45 |
+
# Optional: control (deterministic) forecasts
|
46 |
+
$ ./process.sh ukmo
|
47 |
+
$ ./process.sh ncep
|
48 |
+
$ ./process.sh cma
|
49 |
+
$ ./process.sh ecmwf
|
50 |
+
|
51 |
+
# Optional: perturbed (ensemble) forecasts
|
52 |
+
$ ./process.sh ukmo_ensemble
|
53 |
+
$ ./process.sh ncep_ensemble
|
54 |
+
$ ./process.sh cma_ensemble
|
55 |
+
$ ./process.sh ecmwf_ensemble
|
56 |
```
|
57 |
|
58 |
+
## π Dataset Overview
|
59 |
+
All data has daily and 1.5-degree resolution.
|
60 |
+
|
61 |
+
1. __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:
|
62 |
|
|
|
|
|
|
|
|
|
63 |
Parameters/Levels (hPa) | 1000 | 925 | 850 | 700 | 500 | 300 | 200 | 100 | 50 | 10
|
64 |
:---------------------- | :----| :---| :---| :---| :---| :---| :---| :---| :--| :-|
|
65 |
Geopotential height, z ($gpm$) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
|
|
|
69 |
V component of wind, v ($ms^{-1}$) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
|
70 |
Vertical velocity, w ($Pas^{-1}$) | | | | | ✓ | | | | | |
|
71 |
|
72 |
+
2. __LRA5 Reanalysis__ for Terrestrial (1979-2023)
|
73 |
+
|
74 |
+
| Acronyms | Long Name | Units |
|
75 |
+
|------------------|-------------------------------------------|-----------------------|
|
76 |
+
| asn | snow albedo | (0 - 1) |
|
77 |
+
| d2m | 2-meter dewpoint temperature | K |
|
78 |
+
| e | total evaporation | m of water equivalent |
|
79 |
+
| es | snow evaporation | m of water equivalent |
|
80 |
+
| evabs | evaporation from bare soil | m of water equivalent |
|
81 |
+
| evaow | evaporation from open water | m of water equivalent |
|
82 |
+
| evatc | evaporation from top of canopy | m of water equivalent |
|
83 |
+
| evavt | evaporation from vegetation transpiration | m of water equivalent |
|
84 |
+
| fal | forecaste albedo | (0 - 1) |
|
85 |
+
| lai\_hv | leaf area index, high vegetation | $m^2 m^{-2}$ |
|
86 |
+
| lai\_lv | leaf area index, low vegetation | $m^2 m^{-2}$ |
|
87 |
+
| pev | potential evaporation | m |
|
88 |
+
| ro | runoff | m |
|
89 |
+
| rsn | snow density | $kg m^{-3}$ |
|
90 |
+
| sd | snow depth | m of water equivalent |
|
91 |
+
| sde | snow depth water equivalent | m |
|
92 |
+
| sf | snowfall | m of water equivalent |
|
93 |
+
| skt | skin temperature | K |
|
94 |
+
| slhf | surface latent heat flux | $J m^{-2}$ |
|
95 |
+
| smlt | snowmelt | m of water equivalent |
|
96 |
+
| snowc | snowcover | \% |
|
97 |
+
| sp | surface pressure | Pa |
|
98 |
+
| src | skin reservoir content | m of water equivalent |
|
99 |
+
| sro | surface runoff | m |
|
100 |
+
| sshf | surface sensible heat flux | $J m^{-2}$ |
|
101 |
+
| ssr | net solar radiation | $J m^{-2}$ |
|
102 |
+
| ssrd | download solar radiation | $J m^{-2}$ |
|
103 |
+
| ssro | sub-surface runoff | m |
|
104 |
+
| stl1 | soil temperature level 1 | K |
|
105 |
+
| stl2 | soil temperature level 2 | K |
|
106 |
+
| stl3 | soil temperature level 3 | K |
|
107 |
+
| stl4 | soil temperature level 4 | K |
|
108 |
+
| str | net thermal radiation | $J m^{-2}$ |
|
109 |
+
| strd | downward thermal radiation | $J m^{-2}$ |
|
110 |
+
| swvl1 | volumetric soil water layer 1 | $m^3 m^{-3}$ |
|
111 |
+
| swvl2 | volumetric soil water layer 2 | $m^3 m^{-3}$ |
|
112 |
+
| swvl3 | volumetric soil water layer 3 | $m^3 m^{-3}$ |
|
113 |
+
| swvl4 | volumetric soil water layer 4 | $m^3 m^{-3}$ |
|
114 |
+
| t2m | 2-meter temperature | K |
|
115 |
+
| tp | total precipitation | m |
|
116 |
+
| tsn | temperature of snow layer | K |
|
117 |
+
| u10 | 10-meter u-wind | $ms^{-1}$ |
|
118 |
+
| v10 | 10-meter v-wind | $ms^{-1}$ |
|
119 |
+
|
120 |
+
|
121 |
+
3. __ORAS Reanalysis__ for Sea-Ice (1979-2023)
|
122 |
+
|
123 |
+
| Acronyms | Long Name | Units |
|
124 |
+
|------------------|-------------------------------------------|-----------------------|
|
125 |
+
| iicethic | sea ice thickness | m |
|
126 |
+
| iicevelu | sea ice zonal velocity | $ms^{-1}$ |
|
127 |
+
| iicevelv | sea ice meridional velocity | $ms^{-1}$ |
|
128 |
+
| ileadfra | sea ice concentration | (0-1) |
|
129 |
+
| so14chgt | depth of 14$^\circ$ isotherm | m |
|
130 |
+
| so17chgt | depth of 17$^\circ$ isotherm | m |
|
131 |
+
| so20chgt | depth of 20$^\circ$ isotherm | m |
|
132 |
+
| so26chgt | depth of 26$^\circ$ isotherm | m |
|
133 |
+
| so28chgt | depth of 28$^\circ$ isotherm | m |
|
134 |
+
| sohefldo | net downward heat flux | $W m^{-2}$ |
|
135 |
+
| sohtc300 | heat content at upper 300m | $J m^{-2}$ |
|
136 |
+
| sohtc700 | heat content at upper 700m | $J m^{-2}$ |
|
137 |
+
| sohtcbtm | heat content for total water column | $J m^{-2}$ |
|
138 |
+
| sometauy | meridonial wind stress | $N m^{-2}$ |
|
139 |
+
| somxl010 | mixed layer depth 0.01 | m |
|
140 |
+
| somxl030 | mixed layer depth 0.03 | m |
|
141 |
+
| sosaline | salinity | Practical Salinity Units (PSU) |
|
142 |
+
| sossheig | sea surface height | m |
|
143 |
+
| sosstsst | sea surface temperature | $^\circ C$ |
|
144 |
+
| sowaflup | net upward water flux | $kg/m^2/s$ |
|
145 |
+
| sozotaux | zonal wind stress | $N m^{-2}$ |
|
146 |
+
|
147 |
+
|
148 |
- __Baselines:__
|
149 |
+
- Physics-based models (including control/perturbed forecasts):
|
150 |
- [x] UKMO: UK Meteorological Office
|
151 |
- [x] NCEP: National Centers for Environmental Prediction
|
152 |
- [x] CMA: China Meteorological Administration
|
|
|
158 |
- [x] UNet
|
159 |
- [x] ViT/ClimaX
|
160 |
- [x] PanguWeather
|
|
|
161 |
- [x] GraphCast
|
162 |
+
- [x] Fourcastnetv2
|
163 |
|
164 |
+
## π
Evaluation Metrics
|
165 |
+
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.
|
166 |
|
167 |
+
|
168 |
+
1. __Deterministic-based:__
|
169 |
- [x] RMSE
|
170 |
- [x] Bias
|
171 |
- [x] Anomaly Correlation Coefficient (ACC)
|
172 |
- [x] Multiscale Structural Similarity Index (MS-SSIM)
|
173 |
+
2. __Physics-based:__
|
174 |
- [x] Spectral Divergence (SpecDiv)
|
175 |
- [x] Spectral Residual (SpecRes)
|
176 |
+
|
177 |
+
3. __Probabilistic-based:__
|
178 |
+
- [x] RMSE Ensemble
|
179 |
+
- [x] Bias Ensemble
|
180 |
+
- [x] ACC Ensemble
|
181 |
+
- [x] MS-SSIM Ensemble
|
182 |
+
- [x] SpecDiv Ensemble
|
183 |
+
- [x] SpecRes Ensemble
|
184 |
+
- [x] Continuous Ranked Probability Score (CRPS)
|
185 |
+
- [x] Continuous Ranked Probability Skill Score (CRPSS)
|
186 |
+
- [x] Spread
|
187 |
+
- [x] Spread/Skill Ratio
|
188 |
|
189 |
|
190 |
+
## πͺ Leaderboard
|
191 |
You can access the full score and checkpoints in `logs/<MODEL_NAME>` within the following subdirectory:
|
192 |
- Scores: `eval/<METRIC>.csv`
|
193 |
- Model checkpoints: `lightning_logs/`
|