Source-linked AI summary
Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere
Jonathan A. Weyn, Dale R. Durran, Rich Caruana
TL;DR
The paper addresses whether deep learning can provide accurate, global, purely data-driven weather forecasts while avoiding limitations of latitude–longitude CNN grids and single-step training. It develops a cubed-sphere CNN with improved architecture and multi-step loss, outperforming persistence, climatology, and coarse-resolution NWP in reported comparisons while remaining less accurate than high-resolution operational NWP but much faster.
Problem
The study asks whether deep learning can forecast global weather purely from data and overcome the geographic and long-term-stability limitations of earlier CNN weather models.
Method
The model uses volume-conservative cubed-sphere remapping, a U-Net-based CNN, and multi-time-step loss to forecast several atmospheric variables globally.
Results
The model outperforms persistence, climatology, and coarse-resolution T42 IFS forecasts, but T63 IFS and operational S2S forecasts perform better.
Takeaways & Limitations
DLWP can produce a global four-week forecast in less than two tenths of a second, enabling rapid large-ensemble forecasting.
Takeaways & Limitations
DLWP underestimates the development of blocking patterns and is outperformed by higher-resolution dynamical forecasting systems.
Abstract
from arXiv · showhide
We present a significantly-improved data-driven global weather forecasting framework using a deep convolutional neural network (CNN) to forecast several basic atmospheric variables on a global grid. New developments in this framework include an offline volume-conservative mapping to a cubed-sphere grid, improvements to the CNN architecture, and the minimization of the loss function over multiple steps in a prediction sequence. The cubed-sphere remapping minimizes the distortion on the cube faces on which convolution operations are performed and provides natural boundary conditions for padding in the CNN. Our improved model produces weather forecasts that are indefinitely stable and produce realistic weather patterns at lead times of several weeks and longer. For short- to medium-range forecasting, our model significantly outperforms persistence, climatology, and a coarse-resolution dynamical numerical weather prediction (NWP) model. Unsurprisingly, our forecasts are worse than those from a high-resolution state-of-the-art operational NWP system. Our data-driven model is able to learn to forecast complex surface temperature patterns from few input atmospheric state variables. On annual time scales, our model produces a realistic seasonal cycle driven solely by the prescribed variation in top-of-atmosphere solar forcing. Although it is currently less accurate than operational weather forecasting models, our data-driven CNN executes much faster than those models, suggesting that machine learning could prove to be a valuable tool for large-ensemble forecasting.
1 Introduction
Data-driven weather forecasting has progressed from using machine learning alongside dynamical models toward purely data-driven global prediction. This study extends earlier CNN work with a global cubed-sphere framework and other improvements to address prior geographic and architectural limitations.
- Machine learning has been used to improve NWP outputs, downscale GCM forecasts, estimate forecast uncertainty, and identify or predict extreme weather.
- Purely data-driven weather forecasting became feasible to investigate because of reliable long-term weather observations, open-source machine-learning software, and efficient GPU computing.
- WDC19 used deep CNNs trained on more than 20 years of reanalysis data to forecast northern-hemisphere 500-hPa height and 300–700-hPa thickness.
- WDC19 outperformed climatology for 500-hPa-height RMSE to about 5 days, but used only a northern-hemisphere latitude–longitude grid with inappropriate polar and equatorial boundary conditions.
- This study improves the earlier model with volume-conservative global cubed-sphere mapping, cube-face CNNs, an improved architecture, and sequence prediction.
2 The DLWP model
DLWP maps atmospheric states forward with a fully convolutional CNN adapted to a cubed-sphere representation, using face-aware convolutions, U-Net structure, and multi-step training. The design preserves spatial locality and connects neighboring cube faces while addressing long-horizon instability.
- DLWP uses a fully convolutional neural network to map an initial multidimensional atmospheric state to a future atmospheric state.
- The model combines cubed-sphere remapping, an improved U-Net architecture, and sequence prediction for medium-range and longer forecasts.
- 2.1.1 Description of the grid: The equiangular gnomonic cubed sphere avoids latitude–longitude polar singularities while preserving local interactions for planar convolutions.
- 2.1.1 Description of the grid: Offline forward and inverse remapping uses volume-conservative transformations between latitude–longitude data and six cube faces.
- 2.1.2 Convolution on the cubed sphere: Convolutions operate separately on cube faces, with distinct learned weights for equatorial and polar regions and boundary padding from neighboring faces.
- 2.2 CNN architecture: The U-Net architecture uses convolutional encoding and decoding with skip connections, while the CNN learns local atmospheric features through trainable filters.
- 2.3 Sequence prediction: Multi-time-step training minimizes errors over iterated predictions because single-step optimization can produce short-range skill but unstable long-range forecasts.
3 Data
DLWP uses ERA5 data re-gridded to a global cubed-sphere representation, with distinct training, validation, and test periods. Evaluation compares the model with climatology and multiple IFS benchmarks after harmonizing forecasts onto a common grid.
- Data: ERA5 data from 1979–2018 were retrieved every 3 hours and re-gridded to a cubed sphere with 48 points per cube-face side.The source data were processed on a global 2-degree latitude-longitude grid before cubed-sphere remapping, corresponding to roughly 1.9° spacing near equatorial face centers.
- Data: Training used 1979–2012, validation used 2013–2016, and testing used 2017–2018.The periods were separated to reduce correlations between evaluation and neighboring training times; climatological shifts were assumed negligible.
- Data: The model uses four atmospheric fields: Z500, Z1000, 300–700-hPa geopotential thickness, and 2-meter temperature.The geopotential heights identify mid-latitude weather-system structure, while thickness provides information about weather-system growth and decay.
- Data: Three prescribed fields—insolation, land-sea mask, and topographic height—are inputs but not model outputs.Insolation supplies information about diurnal and annual cycles, especially for predicting 2-meter temperature.
- Benchmarks: Evaluation benchmarks included daily-mean climatology, T42 and T63 IFS models, and the operational subseasonal-to-seasonal IFS.The T42 IFS provides a coarse-resolution dynamical comparison, while the S2S IFS uses 16-km resolution initially and targets 2-week to 2-month forecasts.
- Evaluation: Forecasts and verifying ERA5 data were mapped to the cubed sphere and back to a regular 2° latitude-longitude grid for uniform comparison.DLWP was initialized at the same times as the operational S2S model for direct comparison.
4 Results
DLWP is evaluated against persistence, climatology, and dynamical NWP benchmarks using global error metrics, multi-week forecast states, and a one-year free-running simulation. It outperforms coarse benchmarks at short-to-medium ranges, remains stable at long lead times, and captures broad seasonal behavior, while producing weaker extremes and variability.
- Evaluation framework: DLWP is evaluated with globally averaged RMSE and ACC for Z500 and daily mean T2, plus multi-week state comparisons and a one-year free-running forecast.RMSE measures point-by-point forecast accuracy, while ACC assesses the spatial pattern of anomalies.
- Globally averaged forecast error: 7 days: DLWP beats climatology for Z500 RMSE, while outperforming persistence at all lead times and T42 IFS throughout the available range.T63 IFS and operational S2S perform better; S2S remains below climatology error beyond 9 days.
- Globally averaged forecast error: 9.5 days: operational S2S maintains Z500 ACC above 0.5, compared with just under 7 days for DLWP.DLWP still outperforms persistence and produces spatial patterns with reasonable disturbance amplitudes rather than smooth climatological states.
- Globally averaged forecast error: 8 days: DLWP T2 RMSE remains below climatology, while clearly outperforming persistence and the coarse-resolution T42 IFS.T63 IFS beats climatology to 9 days and S2S to 11 days; early IFS and S2S errors partly reflect initialization or grid differences.
- Multi-week forecast states: 210 four-week forecasts were free from instabilities, enabling arbitrarily long global free-running forecasts without equatorial or polar lateral-boundary errors.The model is a true global model rather than a northern-hemisphere cylindrical-grid system.
- Multi-week forecast states: 642 hours: a selected DLWP forecast reproduced a strong cyclone and associated warm-sector temperature pattern at more than four weeks.The forecast cyclone was weaker than observed, and the match was described as serendipitous because deterministic forecasts have no skill at 26.75 days.
- Annual-cycle behavior: The one-year forecast captures the broad annual Z500 cycle using seasonal top-of-atmosphere insolation and preceding atmospheric states.It reproduces lower polar-winter Z500 followed by spring increases and the next summer onset in approximate agreement with the observed cycle.
- Annual-cycle behavior: The annual simulation has an equatorward 560-dam contour bias after two weeks and weaker temporal variability than observations, while longer forecasts underestimate blocking and extreme patterns.These limitations indicate room for improvement in representing strong ridges, drought- and heat-wave-associated blocking, and longer-term weather variability.
5 Discussion and Conclusions
The extended DLWP model delivers stable global forecasts and outperforms a coarse-resolution dynamical model, while remaining less accurate than higher-resolution operational systems. Its main practical advantage is computational speed.
- The global DLWP model remains stable in free-running one-year forecasts and produces realistic atmospheric states with modest differences from observations.
- It clearly outperforms the coarse-resolution T42 ECMWF IFS in global averaged RMS error and anomaly correlation coefficient.
- The model outperforms T42 IFS forecasts of 500-hPa height and 2-meter temperature despite using only four input variables and three prescribed fields.
- The T63 IFS and very high-resolution operational S2S model outperform DLWP, defining an accuracy gap at higher model resolutions.
- A global four-week forecast takes less than two tenths of a second after training, enabling a 1000-member one-month ensemble in about three minutes.