0 2 0 2

n u J

1 1

] h p -[ Passage withheld — in our stored copy of this paper these characters are set one per token and run backwards, which is what a PDF extractor returns for a table or a caption printed sideways. It is not what the page of the article looks like, so we do not reproduce it. The tags cut from this paper are unaffected. ]

WeatherBench: A benchmark dataset for data-driven weather forecasting

Stephan Rasp1, Peter D. Dueben2, Sebastian Scher3, Jonathan A. Weyn4, Soukayna Mouatadid5, and Nils Thuerey1 1Technical University of Munich, Germany 2European Centre for Medium-range Weather Forecasts, Reading, UK 3Department of Meteorology and Bolin Centre for Climate Research, Stockholm University, Sweden 4Department of Atmospheric Sciences, University of Washington, Seattle, USA 5Department of Computer Science, University of Toronto, Canada

Correspondence: Stephan Rasp (stephan.rasp@tum.de)

Abstract. Data-driven approaches, most prominently deep learning, have become powerful tools for prediction in many domains. A natural question to ask is whether data-driven methods could also be used to predict global weather patterns days

in advance. First studies show promise but the lack of a common dataset and evaluation metrics make inter-comparison between studies difficult. Here we present a benchmark dataset for data-driven medium-range weather forecasting, a topic of

high scientific interest for atmospheric and computer scientists alike. We provide data derived from the ERA5 archive that

has been processed to facilitate the use in machine learning models. We propose simple and clear evaluation metrics which

will enable a direct comparison between different methods. Further, we provide baseline scores from simple linear regression techniques, deep learning models, as well as purely physical forecasting models. The dataset is publicly available at

https://github.com/pangeo-data/WeatherBench and the companion code is reproducible with tutorials for getting started. We

hope that this dataset will accelerate research in data-driven weather forecasting.

1

Introduction

Deep learning, a branch of machine learning based on multi-layered artificial neural networks, has proven to be a powerful tool

for a wide range of tasks, most notably image recognition and natural language processing (LeCun et al., 2015). More recently,

deep learning has also been used in many fields of natural science. Much of the success of deep learning is based on the ability

of neural networks to recognize patterns in high-dimensional spaces. A natural question to ask then is whether deep learning

can also be used to predict future weather patterns.

Currently, weather (and climate) predictions are based on purely physical computer models, in which the governing equations, or our best approximation thereof, of the atmosphere and ocean are solved on a discrete numerical grid (Bauer et al.,

2015). Overall, this approach has been very successful. However, today’s numerical weather prediction (NWP) models still

have shortcoming for many important applications, for example forecasting mesoscale convective systems over Africa (Vogel

et al., 2018). Furthermore, huge amounts of computing power are required, especially for creating probabilistic forecasts which

1

are usually limited to 50 ensemble members or less. For these reasons and the growing popularity of machine learning (ML)

there has been a growing interest to improve and speed up NWP with data-driven approaches.

ML can be applied to weather prediction in many different ways. Two long-standing applications of ML are post-processing

– the correction of statistical biases in the output of physical models – and statistical forecasting – the prediction of variables

not directly output by the physical model. Traditionally, this has been done using simple linear techniques but more recently

modern machine learning approaches like random forests or neural networks have been explored (Gagne et al., 2014; Taillardat

et al., 2016; McGovern et al., 2017; Lagerquist et al., 2017; Rasp and Lerch, 2018). Typically, these approaches target very

specific variables or locations whereas the general evolution of the atmosphere is still predicted by a physical model. Another

application that has recently been explored using ML is nowcasting, which describes the short range (up to 6 hours) prediction

of precipitation by directly extrapolating radar observation without a physical model involved (Shi et al., 2015, 2017; Agrawal

et al., 2019; Grönquist et al., 2020).

Yet another direction for ML research is hybrid modeling, in which a physical model is combined with data-driven components, for example replacing heuristic cloud or radiation parameterizations (Chevallier et al., 1998; Krasnopolsky et al.,

2005; Rasp et al., 2018; Brenowitz and Bretherton, 2018; Yuval and O’Gorman, 2020). The key idea behind these approaches

is to only replace uncertain (e.g. clouds) or computationally expensive (e.g. line-by-line radiation) model components with

machine learning emulators and leave other model components (e.g. large-scale dynamics) untouched. However such hybrid

models also have drawbacks. First, the interaction between physical and machine learning components are poorly understood

and can lead to unexpected instabilities and biases (Brenowitz and Bretherton, 2019). Second, they are difficult to implement

from a technical perspective because one has to interface the machine learning components with complex climate model code,

typically written in Fortran.

Here we focus on purely-data driven prediction of the global atmospheric flow in the medium-range. Specifically, we select

lead times of 3 and 5 days, for which the atmosphere is still reasonably deterministic but also exhibits complex nonlinear

behaviour, such as baroclinic instabilities and tropical cyclogenesis. This forecast range is important from a societal point of

view because it delivers crucial information for disaster preparation, for example for flooding, cold and hot spells or damaging

winds (Lazo et al., 2009). Creating a good medium-range forecast requires understanding complex atmospheric dynamics

and the interplay between several variables across a range of scales. This sets this challenge apart from post-processing and

statistical forecasting, in which the large-scale dynamics are predicted by a physical model, and nowcasting, in which the

considered evolution is univariate and short-term. In other words, this benchmark closely emulates the task performed by

physical NWP models.

There are several motivations for considering a purely-data driven approach. As mentioned above current NWP is computationally expensive and, nevertheless, has low skill for certain applications. If data-driven models were able to learn a more

efficient representation of the underlying dynamical and physical equations, they might enable computationally cheaper forecasts. This can be useful for many applications, for example creating very large ensembles to better estimate the probability

of extreme events. It is also possible that by learning from a diverse set of data sources, data-driven models can outperform

physical models in areas where the latter struggle. While in this benchmark challenge the focus is on upper-level fields of

2

pressure and temperature – for which physical models perform very well – the hope is that the insights gained from this task

can be leveraged for more impactful application. Further, recent research into interpretable machine learning might provide

scientists with new analysis tools (McGovern et al., 2019; Toms et al., 2019). Finally, there is the basic scientific question to

what extent purely data-driven models can learn the underlying dynamics of the atmosphere.

Note also that while this benchmark is framed as a data-driven prediction challenge, the proposed framework can also be

applied to post-processing using the same metrics.

In machine learning research, the data-driven prediction of future states is an active area of research with applications from

language translation (Sutskever et al., 2014), over audio signals (Oord et al., 2016), to numerical simulations (Morton et al.,

2018). In this context, weather forecasts are a particularly challenging task. The behavior is highly complex and non-linear,

but also exhibits some recurring patterns, albeit only on local scales (Hamill and Whitaker, 2006). As such the the proposed

benchmark poses interesting challenges for deep learning algorithms, e.g., to evaluate different architectures (Ronneberger

et al., 2015; He et al., 2015; Huang et al., 2016), regularization methods (Krogh and Hertz, 1992; Srivastava et al., 2014; Xie

et al., 2017) or optimizers (Graves, 2013; Kingma and Ba, 2014).

In the last couple of years, several studies (summarized in Section 2) have pioneered data-driven, global, medium-range

weather prediction. All of them show that there is some potential in this approach but also highlight the need for further

research. In particular, we currently lack a common benchmark challenge to accelerate progress. Benchmark datasets can have

a huge impact because they make different algorithms quantitatively inter-comparable and foster constructive competition,

particularly in a nascent direction of research. Famous examples are the computer vision datasets MNIST (LeCun et al., 1998)

and ImageNet (Russakovsky et al., 2015). Further, well-curated benchmark datasets make it easier for people from different

fields to work on a problem (Ebert-Uphoff et al., 2017).

Here, we propose a benchmark problem for data-driven weather forecasting. We provide a ready-to-use dataset for download

along with specific metrics to compare different approaches. In this paper, we start by reviewing the previous work done on this

topic (Section 2), describe the dataset (Section 3) and the evaluation metrics (Section 4), and provide several baseline models

(Section 5). Finally, we will highlight several promising directions for further research (Section 6) and conclude with a big

picture view (Section 8).

2 Overview of previous work

In this section, we briefly describe the four existing studies on predicting the large-scale atmospheric state in the medium-range

with a focus on the data, methods and evaluation.

2.1 Dueben and Bauer (2018)

In this study, the authors trained a neural network to predict 500 hPa geopotential (Z500; see Section 4 for details on commonly

used fields), and in some experiments 2 meter-temperature, 1 hour ahead. The training data was taken from the ERA5 archive

for the time period from 2010 to 2017 and regridded to a 6 degree latitude-longitude grid. Two neural network variants were

3

Figure 1. Schematic of data-driven weather forecasting. a) Example of direct weather prediction for 5 days lead time. The input to the neural

network are fields on a latitude-longitude grid. The fields can be several levels of the same variable and/or different variables. The goal is to

predict the same fields some time ahead. b) Iterative forecasts are created from data-driven models trained on a shorter lead time, for example

6 hours, which are then iteratively called up to the required forecast lead time.

used, a fully connected neural network and a spatially localized network, similar to a convolutional neural network (CNN).

After training they then created iterative forecasts up to 120 h lead time for 10 month validation period. They compared their

data-driven forecasts to an operational NWP model and the same model run at a spatial resolution comparable to the datadriven method. One interesting detail is that their networks predict the difference from one time step to the next, instead of the

absolute field. To create these iterative forecasts, they use a third-order Adams-Bashford explicit time-stepping scheme. The

CNN predicting only geopotential performed best but was unable to beat the low-resolution physical baseline.

2.2 Scher (2018) and Scher and Messori (2019b)

These two studies addressed the issue of data-driven weather forecasting in a simplified reality setting. Long runs of simplified

General Circulation Models (GMCs) were used as “reality”. Neural networks were trained to predict the model fields several

days ahead. The neural network architecture are CNNs with an encoder-decoder setup. They take as input the instantaneous

3D model fields at one timestep, and output the same model fields at some time later. In Scher (2018), a separate network was

trained for each lead-time up to 14 days. Scher and Messori (2019b) trained only on 1-day forecasts, and constructed longer

forecasts iteratively. Interestingly, networks trained to directly predict a certain forecast time, e.g. 5 days, outperformed iterative

networks. The forecasts were evaluated using the root mean squared error and the anomaly correlation coefficient of Z500 and

4

Longitudet = 0t = 5 daysa) Direct predictionb) Iterative predictiont = 0t = 6 hourst = 5 daysLatitudeChannels = Variables xLevels 800 hPa temperature. Scher (2018) used a highly simplified GCM without hydrological cycle, and achieved very high predictive

skill. Additionally, they were able to create stable "climate" runs (long series of consecutive forecasts) with the network. Scher

and Messori (2019b) used several more realistic and complex GCMs. The data-driven model achieved relatively good shortterm forecast skill, but was unable to generate stable and realistic “climate” runs. In terms of neural-network architectures they

showed that architectures tuned on simplified GCMs also work on more complex GCMs, and that the same architecture also

has some prediction skill on single-level reanalysis data.

2.3 Weyn et al. (2019)

In this study, reanalysis-derived Z500 and 700-300 hPa thickness at 6-hourly time steps are predicted with deep CNNs. The data

are from the Climate Forecast System (CFS) Reanalysis from 1979–2010 with 2.5-degree horizontal resolution and cropped to

the northern hemisphere. The authors used similar encoder-decoder convolutional networks as those used by Scher (2018) and

Scher and Messori (2019b) but also experimented with adding a convolution long short-term memory (LSTM; Hochreiter and

Schmidhuber, 1997) hidden layer. As in Scher and Messori (2019b), forecasts are generated iteratively by feeding the model’s

outputs back in as inputs. The authors found that using two input time steps, 6 h apart, and predicting two output time steps,

performed better than using a single step. Their best CNN forecast outperforms a climatology benchmark at up to 120 h lead

time, and appears to correctly asymptote towards persistence forecasts at longer lead times up to 14 days.

These three approaches outline promising first steps towards data-driven forecasting. The differences of the proposed methods already highlight the importance of a common benchmark case to compare prediction skill.

3 Dataset

For the proposed benchmark, we use the ERA5 reanalysis dataset (Hersbach et al., 2020) for training and testing. Reanalysis

datasets provide the best guess of the atmospheric state at any point in time by combining a forecast model with the available

observations. The raw data is available hourly for 40 years from 1979 to 2018 on a 0.25°latitude-longitude grid (721×1440

grid points) with 37 vertical levels.

Since this raw dataset is very large (a single vertical level for the entire time period amounts to almost 700GB of data), we

regrid the data to lower resolutions. This is also a more realistic use case, since very high resolutions are still hard to handle for

deep learning models because of GPU memory constraints and I/O speed. In particular, we chose 5.625° (32×64 grid points),

2.8125° (64×128 grid points) and 1.40525° (128×256 grid points) resolution for our data. The regridding was done with the

xesmf Python package (Zhuang, 2019) using a bilinear interpolation. Powers of two for the grid are used since this is common

for many deep learning architectures where image sizes are halved in the algorithm. Further, for 3D fields we selected 13

vertical levels: 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. Note that it is common to use pressure

in hecto-Pascals as a vertical coordinate instead of physical height. The pressure at sea level is approximately 1000 hPa and

decreases roughly exponentially with height. 850 hPa is at around 1.5 km height. 500 hPa is at around 5.5 km height. If the

surface pressure is smaller than a given pressure level, for example at high altitudes, the pressure-level values are interpolated.

5

The selected pressure levels contain the seven pressure levels that are commonly used for 3D output by the climate models in

the Coupled Model Intercomparison Project Phase 6 (CMIP6, Eyring et al., 2016) which could be useful for pretraining. One

regridded historical climate run is also available from the data repository with a template workflow for downloading further

CMIP data on the Github repository.

The processed data (see Table 1) are available at https://mediatum.ub.tum.de/1524895 (Rasp et al., 2020). The data are split

into yearly NetCDF files for each variable and resolution, packed in a zip file. The entire dataset at 5.625° resolution has a size

of 191GB. Individual variables amount to around 25GB three-dimensional and 2GB for two-dimensional fields. File sizes for

2.8125° and 1.40525° resolutions are a factor 4 and 16 times larger. Data processing was organized using Snakemake (Koster and Rahmann, 2012). For further instructions on data downloading visit the Github page1. The available variables were chosen

based on meteorological consideration. Geopotential, temperature, humidity and wind are prognostic state variables in most physical NWP and climate models. Geopotential at a certain pressure level p, typically denoted as Φ with units of m2s−2,

defined as

z at p (cid:90)

g dz(cid:48)

Φ =

0

(1)

where z describes height in meters and g = 9.81 m s−2 is the gravitational acceleration. Horizontal relative vorticity, defined

as ∂v/∂x − ∂u/∂y, describes the rotation of air at a given point in space. Potential vorticity (Hoskins et al., 1985; Holton,

2004) is a commonly used quantity in synoptic meteorology which combines the rotation (vorticity) and vertical temperature gradient of the atmosphere. It is defined as P V = ρ−1ζa · ∇θ, where ρ is the density, ζa is the absolute vorticity (relative plus the Earth’s rotation) and θ is the potential temperature. In addition to the three-dimensional fields, we also include several twodimensional fields: 2 meter-temperature is often used as an impact variable because of its relevance for human activities and

is directly affected by the diurnal solar cycle; 10 meter-wind is also an important impact-related forecast variable, for example

for wind energy; similarly, total cloud cover is an essential variable for solar energy forecasting. We also included precipitation

but urge caution since precipitation in reanalysis datasets often shows large deviation from observations (e.g. Betts et al.,

2019; Xu et al., 2019). Finally, we added the top-of-atmosphere incoming solar radiation as it could be a useful input variable

to encode the diurnal cycle. Further, there are several potentially important time-invariant fields, which are contained in the

constants file. The first three variables enclose information about the surface: the land-sea mask is a binary field with ones for land points; the soil type consists of seven different soil categories2; orography is simply the surface height. In addition,

we included two-dimensional fields with the latitude and longitude values at each point. Particularly the latitude values could

become important for the network to learn latitude-specific information such as the grid structure or the Coriolis effect (see

Section 6). The Github code repository includes all scripts for downloading and processing of the data. This enables users to

download additional variables or regrid the data to a different resolution.

1https://github.com/pangeo-data/WeatherBench 2Coarse = 1, Medium = 2, Medium fine = 3, Fine = 4, Very fine = 5, Organic = 6, Tropical organic = 7, see https://apps.ecmwf.int/codes/grib/param-db?

id=43

6

Table 1. List of variables contained in the benchmark dataset.

Short name Description

z

t

q

r

u

v

vo

pv

t2m

u10

v10

tcc

tp

tisr

lsm

slt

Proportional to the height of a pressure level

Temperature

Mixing ratio of water vapor

Humidity relative to saturation

Wind in x/longitude-direction

Wind in y/latitude direction

Relative horizontal vorticity

Potential vorticity

Temperature at 2 m height above surface

Wind in x/longitude-direction at 10 m height

Wind in y/latitude-direction at 10 m height

Fractional cloud cover

Hourly precipitation

Accumulated hourly incident solar radiation

File containing time-invariant fields

Land-sea binary mask

Soil-type categories

orography

Height of surface

lat2d

lon2d

2D field with latitude at every grid point

2D field with longitude at every grid point

Unit

[m2s−2]

[K] [kg kg−1]

[%] [m s−1] [m s−1] [1 s−1] [K m2 kg−1 s−1]

[K] [m s−1] [m s−1]

(0–1)

[m] [J m−2]

(0/1)

see text

[m]

[°]

[°]

Levels

13 levels

13 levels

13 levels

13 levels

13 levels

13 levels

13 levels

13 levels

Single level

Single level

Single level

Single level

Single level

Single level

Single level

Single level

Single level

Single level

Single level

Long name

geopotential

temperature

specific_humidity

relative_humidity

u_component_of_wind

v_component_of_wind

vorticity

potential_vorticity

2m_temperature

10m_u_component_of_wind

10m_v_component_of_wind

total_cloud_cover

total_precipitation

toa_incident_solar_radiation

constants

land_binary_mask

soil_type

orography

latitude

longitude

4 Evaluation

Evaluation is done for the years 2017 and 2018. To make sure no overlap exists between the training and test dataset, the

first test date is 1 January 2017 00UTC plus forecast time (i.e. for a three day forecast the first test date would be 4 January

2017 00UTC) while the last training target is 31 December 2016 23UTC. Further, the evaluation presented here is done on 5.625° resolution3. This means that predictions at higher resolutions have to be downscaled to the evaluation resolution. We

also evaluated some baselines at higher resolutions and found that the scores were almost identical with differences smaller

than 1%. Therefore we are reassured that little information is lost by evaluating at a coarser resolution.

A note on validation and testing: In machine learning it is good practice to split the data into three parts: the training, validation and test sets. The training dataset is used to actually fit the model. The validation dataset is used during experimentation

3The evaluation of all baselines in this paper are done in this Jupyter notebook: https://github.com/pangeo-data/WeatherBench/blob/master/notebooks/

4-evaluation.ipynb

7

to check the model performance on data not seen during training. However, there is the danger that through continued tuning

of hyperparameters one unwillingly overfits to the validation dataset. Therefore it is advisable to keep a third, testing, dataset

for final evaluations of model performance. For this benchmark this final evaluation is done for the years 2017 and 2018.

Therefore, we strongly encourage users of this dataset to pick a period from 1979 to 2016 for validation of their models for hyperparameter tuning. Because meteorological fields are highly correlated in time, it is advisable to choose a longer contiguous

chunk of data for validation instead of a completely random split. Here we chose the year 2016 for validation.

We chose 500 hPa geopotential and 850 hPa temperature as primary verification fields. Geopotential at 500 hPa pressure,

often abbreviated as Z500, is a commonly used variable that encodes the synoptic-scale pressure distribution. It is the standard

verification variable for most medium-range NWP models. Note that geopotential height, also commonly used, is defined as

Φ/g with units of meters. We picked 850 hPa temperature as our secondary verification field because temperature is a more

impact-related variable. 850 hPa is usually above the planetary boundary layer and therefore not affected by diurnal variations

but provides information about broader temperature trends, including cold spells and heat waves. In addition we also provide

some baseline scores for total 6-hourly accumulated precipitation (TP; but noting the dubious quality mentioned above) and

2-meter temperature (T2M). However, they will not be further discussed here.

We chose the root mean squared error (RMSE) as our primary metric because it is easy to compute and mirrors the loss used

for most ML applications. We define the RMSE as the mean latitude-weighted RMSE over all forecasts:

RMSE =

1 Nforecasts

Nforecasts(cid:88)

(cid:118) (cid:117) (cid:117) (cid:116)

i

1 NlatNlon

Nlat(cid:88)

Nlon(cid:88)

j

k

L(j)(fi,j,k − ti,j,k)2

(2)

where f is the model forecast and t is the ERA5 truth. L(j) is the latitude weighting factor for the latitude at the jth latitude

index:

L(j) =

cos(lat(j)) (cid:80)Nlat j

cos(lat(j))

1 Nlat

(3)

In addition, we also evaluate the baselines using the latitude weighted anomaly correlation coefficient (ACC; see Section

7.6.4 of Wilks, 2006) and the mean absolute error (MAE). The tables and figures can be found in the Appendix A. For smooth

fields like Z500 and T850 the qualitative differences between the metrics are small. For intermittent fields like precipitation the

choice of metric matters a lot more.

5 Baselines

To evaluate the skill of a forecasting model it is important to have baselines to compare to. In this section, we compute scores

for several baselines. The results are summarized in Fig. A3 and Table 2.

5.1 Persistence and Climatology

The two simplest possible forecasts are a) a persistence forecast in which the fields at initialization time are used as forecasts

("tomorrow’s weather is today’s weather"), and b) a climatological forecast. For the climatological forecast, two different

8

Figure 2. RMSE of a) 500 hPa geopotential and b) 850 hPa temperature for different baselines at 5.625° resolution. Solid lines for linear

regression and CNN indicate iterative forecasts, while dots represent direct forecasts for 3 and 5 days lead time.

Table 2. Baseline RMSE for 3 and 5 days forecast time at 5.625° resolution. Best machine learning baseline and physical model are highlighted. TP is 6 hourly accumulated precipitation.

Baseline

Persistence

Climatology

Weekly climatology

RMSE (3 days / 5 days)

Z500 [m2 s−2]

T850 [K]

T2M [K]

TP [mm]

936 / 1033

4.23 / 4.56

3.00 / 3.27

3.23 / 3.24

1075

816

5.51

3.50

6.07

3.19

2.36

2.32

Linear regression (direct)

693 / 783

3.19 / 3.44

2.39 / 2.60

2.37 / 2.37

Linear regression (iterative)

718 / 810

626 / 757

3.17 / 3.48

2.87 / 3.37

CNN (direct)

CNN (iterative)

IFS T42

IFS T63

Operational IFS

1114 / 1559

4.48 / 9.69

489 / 743

268 / 463

154 / 334

3.09 / 3.83

3.21 / 3.69

1.85 / 2.52

2.04 / 2.44

1.36 / 2.03

1.35 / 1.77

2.36 / 2.59

climatologies were computed from the training dataset (1979–2016): first, a single mean over all times in the training dataset

and, second, a mean computed for each of the 52 calendar weeks. The weekly climatology is significantly better, approximately

matching the persistence forecast between 1 and 2 days, since it takes into account the seasonal cycle. This means that to be

useful, a forecast system needs to beat the weekly climatology and the persistence forecast.

9

012345Forecast time [days]020040060080010001200Z500 RMSE [m2 s2]a) Z500PersistenceClimatologyWeekly clim.OperationalIFS T42IFS T63LR (iterative)CNN (iterative)LR (direct)CNN (direct)012345Forecast time [days]0123456T850 RMSE [K]b) T850 5.2 Operational NWP model

The gold standard of medium-range NWP is the operational IFS (Integrated Forecast System) model of the European Center for Medium-range Weather Forecasting (ECMWF)4. We downloaded the forecasts for 2017 and 2018 from the THORPEX Interactive Grand Global Ensemble (TIGGE; Bougeault et al., 2010) archive5, which contains the operational forecasts, initialized at 00 and 12 UTC regridded to a 0.5° by 0.5° grid, which we further regridded to 5.625°. Note that the forecast error starts

above zero because the operational IFS is initialized from a different analysis. Operational forecasting is computationally very

expensive. The current IFS deterministic forecast is computed on a cluster with 11,664 cores. One 10 day forecast at 10 km

resolution takes around 1 hour of real time to compute.

5.3 Physical NWP model run at coarser resolution

To provide physical baselines more in line with the computational resources of a data-driven model, we ran the IFS model

at two coarser horizontal resolutions, T42 (approximately 2.8° or 310 km resolution at the equator (NCAR)) with 62 vertical

levels and T63 (approximately 1.9° or 210 km) with 137 vertical levels. The T42 run was initialized from ERA5 whereas

the T63 run was initialized from the operational analysis. The gap in skill at t = 0 is caused by the conversion to spherical

coordinates at coarse resolutions. For Z500 the skill for these two runs lies in-between the operational IFS and the machine

learning baselines. For T850, the T42 run is significantly worse. The likely reason for this is that temperature close to the

ground is much more affected by the resolution and representation of topography within the model. Further, the model was

not specifically tuned for these resolutions. Computationally, a single forecast takes 270 seconds for the T42 model and 503

seconds for the T64 model on a single XC40 node with 36 cores. Since the computational costs and resolutions of these runs

are much closer to those of a data-driven method, beating those baselines should be a realistic target. note, however, that the

model was not tuned to run at such coarse resolutions.

5.4 Linear regression

As a first purely data-driven baseline we fit a simple linear regression model. For the direct predictions a separate model was

trained for each of the four variables. For this purpose the 2D fields were flattened from 32×64 → 2048. This was done for 3 d

and 5 d forecast time. In addition an iterative model for Z500 and T850 was trained. Here we use a single linear regression to

predict 6 hours ahead where the two fields are concatenated (2×32×64 → 4096). The advantage of iterative forecasts is that a

single model is able to make predictions for any forecast time rather than having to train several models. For iterative forecasts

the model takes its previous output as input for the next step. To create a 5 day iterative forecast the model trained to predict

6 hour forecasts is called 20 times. For this model, the iterative forecast performs just as well as the direct forecast due to its

linear nature. At 5 days, the linear regression forecast is about as good as the weekly climatology.

4https://www.ecmwf.int/en/forecasts/documentation-and-support 5The TIGGE data for total precipitation and 2m temperature was damaged for 2017. For this reason the TIGGE evaluation for these variables is only done

using the 2018 data.

10

5.5 Simple convolutional neural network

As our deep learning baseline we chose a simple fully-convolutional neural network. CNNs are the natural choice for spatial

data since they exploit translational invariances in images/fields. Here we train a CNN with 5 layers. Each hidden layer has

64 channels with a convolutional kernel of size 5 and ELU activations (Clevert et al., 2015). The input and output layers have

two channels, representing Z500 and T850. The model was trained using the Adam optimizer (Kingma and Ba, 2014) and a

mean squared error loss function. The total number of trainable parameters is 313,858. We implemented periodic convolutions

in the longitude direction but not the latitude direction. The implementation can be found in the Github repository. The direct

CNN forecasts beat the linear regression forecasts for 3 and 5 days forecast time. However, at 5 days these forecasts are only

marginally better than the weekly climatology (see Table 2). This baseline, however, is to be seen simply as a starting point

for more sophisticated data driven methods. The iterative CNN forecast, which equivalently to the linear regression iterative

forecast was created by chaining together 6 hourly predictions, performs well up to around 1.5 days but then the network’s

errors grow quickly and diverge. This confirms the findings of Scher and Messori (2019a) whose experiments showed that

training with longer lead time yields better results than chaining together short-term forecasts. However, the poor skill of the

iterative forecast could easily be a result of using an overly simplistic network architecture. The iterative forecasts of Weyn

et al. (2019), who employ a more complex network structure, show stable long term performance up to two weeks with realistic

statistics.

5.6 Example forecasts

To further illustrate the prediction task, Fig. 3 shows example geopotential and temperature fields. The ERA5 temporal differences show several interesting features. First, the geopotential fields and differences are much smoother compared to the

temperature fields. The differences in both fields are also much smaller in the tropics compared to the extratropics where propagating fronts can cause rapid temperature changes. An interesting feature is detectable in the 6h Z500 difference field in the

tropics. These alternating patterns are signatures of atmospheric tides.

The CNN forecasts for 6h lead time are not able to capture these wave-like patterns which hints at a failure to capture the

basic physics of the atmosphere. For 5 days forecast time the CNN model predicts unrealistically smooth fields. This is likely

a result of two factors: first, the two input fields used in this baseline CNN contain insufficient information to create a skillful

5 day forecast; and second, at 5 days the atmosphere already shows some chaotic behavior which causes a model trained with

a simple RMSE loss to predict smooth fields (see Section 6). The IFS operational forecast has much smaller errors than the

CNN forecast. It is able to capture the propagation of tropical waves. Its main errors appear at 5 days in the mid-latitudes where

extratropical cycles are in slightly wrong positions.

11

Figure 3. Example fields for 2017-01-01 00UTC initialization time. The top two rows show the ERA5 "truth" fields for geopotential (Z500)

and temperature (T850) at initialization time (t=0h) and for 6h and 5d forecast time. In addition, the difference between the forecast times

and the initialization time is shown. The third and fourth rows show the forecasts from the CNN model. Rows five and six show the IFS

operational model. For the CNN forecasts the first column is identical to the ERA5 truth. We selected the 6h iterative CNN model for the 6h

forecast but the 5d direct CNN model for the 5 day forecast. For the IFS the initial states (t=0h) differ slightly albeit not visibly. In addition

to the forecast fields the error relative to the ERA5 "truth" is shown in the third and fifth columns. Please note that the colorbars for the

difference fields change.

12

ERA5 “Truth”CNN forecastsIFS forecasts 6 Discussion

6.1 Weather-specific challenges

From a ML perspective, state-to-state weather prediction is similar to image-to-image translation. For this sort of problem many

deep learning techniques have been developed in recent years (Kaji and Kida, 2019). However, forecasting weather differs in

some important ways from typical image-to-image applications and raises several open questions.

First, the atmosphere is three-dimensional. So far, this aspect has not been taken into account. In the networks of Scher and

Messori (2019a), for example, the different levels have been treated as separate channels of the CNN. However, simply using a

three-dimensional CNN might not work either because atmospheric dynamics and grid spacings change in the vertical, thereby

violating the assumption of translation invariance which underlies the effectiveness of CNNs. This directly leads to the next

challenge: On a regular latitude-longitude grid, the dynamics also change with latitude because towards the poles the grid cells

become increasingly stretched. This is in addition to the Coriolis effect, the deflection of wind caused by the rotation of Earth,

which also depends on latitude. A possible solution in the horizontal could be to use spherical convolutions (Cohen et al., 2018;

Perraudin et al., 2019; Jiang et al., 2019) or to feed in latitude information to the network.

Another potential issue is the limited amount of training data available. 40 years of hourly data amounts to around 350,000

samples. However, the samples are correlated in time. If one assumes that a new weather situation occurs every day, then

the number of samples is reduced to around 15,000. Without empirical evidence it is hard to estimate whether this number

is sufficient to train complex networks without overfitting. Should overfitting be a problem, one could try transfer learning.

In transfer learning, the network is pretrained on a similar task or dataset, for example, climate model simulations, and then

finetuned on the actual data. This is common practice in computer vision and has been successfully applied to seasonal ENSO

forecasting (Ham et al., 2019). Another common method to prevent overfitting is data augmentation, which in traditional

computer vision is done by e.g. randomly rotating or flipping the image. However, many of the traditional data augmentation

techniques are questionable for physical fields. Random rotations, for example, will likely not work for this dataset since the x

and y directions are physically distinct. Thus, finding good data augmentation techniques for physical fields is an outstanding

problem. Using ensemble analyses and forecasts could provide more diversity in the training dataset.

Finally, there are technical challenges. Data for a single variable with ten levels at 5.625° resolution take up around 30 GB of

data. For a network with several variables or even at higher resolution, the data might not fit into CPU RAM any more and data loading could become a bottleneck. For image files, efficient data loaders have been created6. For netCDF files, however, so

far no efficient solution exists to our knowledge. Further, one can assume that to create a competitive data-driven NWP model,

high resolutions have to be used, for which GPU RAM quickly becomes a limitation. This suggests that multi-GPU training

might be necessary to scale up this approach (potentially similar to the technical achievement of Kurth et al. (2018)).

6See e.g. https://keras.io/preprocessing/image/ or https://pytorch.org/tutorials/beginner/data_loading_tutorial.html. One promising but so far unexplored

option is to use Tensorflow’s TFRecords (https://www.tensorflow.org/tutorials/load_data/tfrecord)

13

6.2 Probabilistic forecasts and extremes

One important aspect that is not currently addressed by this benchmark is probabilistic forecasting. Because of the chaotic

evolution of the atmosphere, it is very important to also have an estimate of the uncertainty of a forecast. In physical NWP

this is done by running several forecasts, called an ensemble, from slightly different initial conditions and potentially with

different or stochastic model physics (Palmer, 2019). From this Monte Carlo forecast one can then estimate a probability

distribution. A different approach, which is often taken in statistical post-processing of NWP forecasts, is to directly estimate a

parametric distribution (e.g. Gneiting et al., 2005; Rasp and Lerch, 2018). For a probabilistic forecast to be reliable the forecast

uncertainty has to be an accurate indicator of the error. A good first order approximation for this is the spread (ensemble

standard deviation) to error (RMSE) ratio which should be one (Leutbecher and Palmer, 2008). A more stringent test is to

use a proper probabilistic scoring rule, for example the continuous ranked probability score (CRPS) (Gneiting and Raftery,

2007; Gneiting and Katzfuss, 2014). For deterministic forecast the CRPS reduces to the mean absolute error. Extending this

benchmark to probabilistic forecasting simply requires computing a probabilistic score. How to produce probabilistic datadriven forecasts is a very interesting research question in its own right. We encourage users of this benchmark to explore this

dimension.

A related issue is the question of extreme weather situations, for example heat waves. These events are, by definition, rare,

which means that they will contribute little to regular verification metrics like the RMSE. However, for society these events

are highly important. For this reason, it would make sense to evaluate extreme situations separately. But defining extremes

is ambiguous which is why there is no standard metric for evaluating extremes. The goal of this benchmark is to provide a

simple, clear problem. Therefore, we decided to omit extremes for now but users are encouraged to chose their own verification

of extremes.

6.3 Climate simulations

Another aspect that is untouched by the benchmark challenge proposed here is climate prediction. Even though weather and

climate deal with the same underlying physical system, they pose different forecasting challenges. In weather forecasting the

goal is to predict the state of the atmosphere at a specific time into the future. This is only possible up to the prediction

horizon of the atmosphere, which is thought to be at roughly two weeks. Climate models, on the other hand, are evaluated

by comparing long-term statistics to observations, for example the mean surface temperature (Stocker et al., 2013). Scher and

Messori (2019a) created iterative climate time scale runs with their data-driven models and compared first and second-order

statistics. They found that the model sometimes produced a stable climate but with significant biases and a poor seasonal

cycle. This indicates that, so far, iterative data-driven models have been unable to produce physically reasonable long-term

predictions. This remains a key challenge for future research. While not specifically included in this benchmark, a good test for

climate simulations is to look at long term mean statistics and the seasonal cycle as done in Figs. 6 and 7 of Scher and Messori

(2019a).

14

Climate change simulations represent another step up in complexity. To start with, external greenhouse gas forcing would

have to be included. Further, future climates will produce atmospheric states that lie outside of the historical manifold of states.

Plain neural networks are very bad at extrapolating to climates beyond what they have seen in the training dataset (Rasp et al.,

2018). For this reason, climate change simulations with current data-driven methods are likely not a good idea. However,

research into physical machine learning is ongoing and might offer new opportunities in the near future (e.g. Bar-Sinai et al.,

2019; Beucler et al., 2019).

6.4 Promising research directions

There is a wide variety of promising research directions for data-driven weather forecasting. The most obvious direction is

to increase the amount of data used for training and the complexity of the network architecture. This dataset provides a, so

far, unexploited volume and diversity of data for training. It is up to future research to find out exactly which combination

of variables will turn out to be useful. Further, this dataset offers a four times higher horizontal resolution than all previous

studies. The hope is that this data will enable researcher to train more complex models than have previously been used.

With regards to model architecture, there is a huge variety of network architectures that can be explored. U-Nets (Ronneberger et al., 2015) have been used extensively for image segmentation tasks that require computations across several spatial

scales. Resnets (He et al., 2015) are currently the state of the art for image classification and their residual nature could be a

good fit for state-to-state forecasting tasks. For synthesis tasks, generative adverserial networks (GANs) (Goodfellow et al.,

2014) were shown to be particularly powerful for creating realistic natural images and fluid flows (Xie et al., 2018). This might

be attractive since minimizing a mean loss, such as the MSE, for random or stochastic data leads to unrealistically smooth

predictions as seen in Fig. 3. Conditional GANs (Mirza and Osindero, 2014; Isola et al., 2016) could potentially alleviate this

issue but it is still unclear to what extent GAN predictions are able to recover the multi-variate distribution of the training

samples.

7 Code and data availability

The dataset is available at https://mediatum.ub.tum.de/1524895 (Rasp et al., 2020). Code, instructions for dowloading the data

and evaluating forecasts can be found at https://github.com/pangeo-data/WeatherBench.

8 Conclusions

In this paper a benchmark dataset for data-driven weather forecasting is presented. It focuses on global medium-range (roughly

2 days to 2 weeks) prediction. With the rise of deep learning in physics, weather prediction is a challenging and interesting

target because of the large overlap with traditional deep learning tasks (Reichstein et al., 2019). While first attempts have

been made in this direction, as discussed in Section 2, the field currently lacks a common dataset which enables the inter-

15

comparison of different methods. We hope that this benchmark can provide a foundation for accelerated research in this area.

Loosely following Ebert-Uphoff et al. (2017), the key features of this benchmark are:

– Scientific impact: Numerical weather forecasting impacts many aspects of society. Currently, NWP model run on massive super-computers at very high computational cost. Building a capable data-driven model would be beneficial in many

ways (see Section 1). In addition, there is the open, and highly debated, question whether fully data-driven methods are

able to learn a good representation of atmospheric physics.

– Challenge for data science: While global weather prediction is conceptually similar an image-to-image task, and therefore allows for the application of many state-of-the-art deep learning techniques, there are some unique challenges to this

problem: the three-dimensional, anisotropic nature of the atmosphere; non-uniform-grids; potentially limited amounts of

training data and the technical challenge of handling large data volumes.

– Clear metric for success: We defined a single metric (RMSE) for two fields (500 hPa geopotential and 850 hPa temperature). These scores provide a simple measure of success for data-driven, medium-range forecast model.

– Quick start: The code repository contains a quick-start Jupyter notebook for reading the data, training a neural network

and evaluating the predictions against the target data. In addition, the repository contains many functions which are likely

to be used frequently, for example an implementation of periodic convolutions in Keras.

– Reproducibility and citability: All baselines and results form this paper are fully reproducible from the code repository.

Further, the baseline predictions are all saved in the data repository. The data has been assigned a permanent DOI.

– Communication platform: We will use the Github code repository as an evolving hub for this project. We encourage

users of this dataset to start by forking the repository and eventually merge code that might be useful for others back into

the main branch. The main platform for communication, e.g. asking questions, about this project will be Github issues.

We hope that this benchmark will foster collaboration between atmospheric and data scientists in the ways we imagined and

beyond.

16

(A1)

(cid:80) tj,k.

Appendix A: Additional metrics

The anomaly correlation coefficient (ACC) is defined as

ACC =

(cid:80)

i,j,kt(cid:48) i,j,k L(j)f (cid:48) (cid:80)

i,j,k L(j)f (cid:48)2

i,j,k

i,j,k i,j,k L(j)t(cid:48)2

i,j,k

(cid:113)(cid:80)

where the prime (cid:48) denotes the difference to the climatology. Here the climatology is defined as climatologyj,k = 1 The mean absolute error is defined just like the MSE (Eq. 2) but with the absolute instead of the squared difference.

Ntime

Figure A1. RMSE of a) 2-meter temperature and b) 6-hourly accumulated precipitation for different baselines at 5.625° resolution.

17

012345Forecast time [days]0123456T2M RMSE [K]a) T2MPersistenceClimatologyWeekly clim.OperationalIFS T42IFS T63LR (direct)012345Forecast time [days]0.00.51.01.52.02.53.03.5TP RMSE [mm]b) TP Figure A2. ACC of a) 500 hPa geopotential, b) 850 hPa temperature, c) 2-meter temperature and d) 6-hourly accumulated precipitation for

different baselines at 5.625° resolution.

18

012345Forecast time [days]0.50.60.70.80.91.0Z500 ACCa) Z500PersistenceClimatologyWeekly clim.OperationalIFS T42IFS T63LR (iterative)CNN (iterative)LR (direct)CNN (direct)012345Forecast time [days]0.600.650.700.750.800.850.900.951.00T850 ACCb) T850012345Forecast time [days]0.40.50.60.70.80.91.0T2M ACCc) T2M012345Forecast time [days]0.00.20.40.60.81.0TP ACCd) TP Figure A3. MAE of a) 500 hPa geopotential, b) 850 hPa temperature, c) 2-meter temperature and d) 6-hourly accumulated precipitation for

different baselines at 5.625° resolution.

19

012345Forecast time [days]0100200300400500600700Z500 MAE [m2 s2]a) Z500PersistenceClimatologyWeekly clim.OperationalIFS T42IFS T63LR (iterative)CNN (iterative)LR (direct)CNN (direct)012345Forecast time [days]0.00.51.01.52.02.53.03.54.0T850 MAE [K]b) T850012345Forecast time [days]0.00.51.01.52.02.53.03.5T2M MAE [K]a) T2M012345Forecast time [days]0.00.20.40.60.81.0TP MAE [mm]b) TP Table A1. Baseline ACC for 3 and 5 days forecast time at 5.625° resolution. Best machine learning baseline and physical model are highlighted. TP is 6 hourly accumulated precipitation.

Baseline

Persistence

Climatology

Weekly climatology

ACC (3 days / 5 days)

Z500

T850

T2M

TP

0.62 / 0.53

0.69 / 0.65

0.88 / 0.85

0.06/0.06

0

0.65

0

0.77

0

0.85

0

0.16

Linear regression (direct)

0.76 / 0.68

0.81 / 0.78

0.92 / 0.90

0.15 / 0.13

Linear regression (iterative)

0.76 / 0.67

0.82 / 0.78

CNN (direct)

CNN (iterative)

IFS T42

IFS T63

0.81 / 0.71

0.85 / 0.79

0.61 / 0.41

0.72 / 0.31

0.90 / 0.78

0.86 / 0.78

0.87 / 0.83

0.97 / 0.91

0.94 / 0.90

0.94 / 0.92

Operational IFS

0.99 / 0.95

0.97 / 0.93

0.98 / 0.96

0.43 / 0.30

Table A2. Baseline MAE for 3 and 5 days forecast time at 5.625° resolution. Best machine learning baseline and physical model are

highlighted. TP is 6 hourly accumulated precipitation.

Baseline

Persistence

Climatology

Weekly climatology

MAE (3 days / 5 days)

Z500 [m2 s−2]

T850 [K]

T2M [K]

TP [mm]

572 / 634

2.91 / 3.12

1.81 / 1.95

1.07 / 1.08

708

525

3.84

2.48

3.77

2.00

0.92

0.88

Linear regression (direct)

431 / 489

2.23 / 2.41

1.48 / 1.60

0.96 / 0.97

Linear regression (iterative)

447 / 508

403 / 476

2.20 / 2.42

2.02 / 2.38

CNN (direct)

CNN (iterative)

IFS T42

IFS T63

Operational IFS

892 / 1263

3.49 / 7.49

295 / 449

188 / 307

97 / 198

1.99 / 2.53

1.83 / 2.12

1.30 / 1.73

1.28 / 1.53

0.93 / 1.35

0.84 / 1.08

0.69 / 0.81

Author contributions. SR, PD, SS and JW conceived the idea. SR prepared the data and baselines and led the writing. All authors contributed

to the manuscript.

20

Competing interests. The authors declare no competing interests.

Acknowledgements. Stephan Rasp acknowledges funding from the German Research Foundation (DFG). We thank the Copernicus Climate

Change Service (C3S) for allowing us to redistribute the data. Peter D. Dueben gratefully acknowledges funding from the Royal Society for

his University Research Fellowship and the ESIWACE2 project. The ESIWACE2 project have received funding from the European Union’s

Horizon 2020 research and innovation programme under grant agreement No 823988.

21