Explore the vast range of titles available.

The data‐driven approaches for medium‐range weather forecasting are recently shown to be extraordinarily promising for ensemble forecasting due to their fast inference speed compared to the traditional numerical weather prediction models. However, their forecast accuracy can hardly match the state‐of‐the‐art operational ECMWF Integrated Forecasting System (IFS) model. Previous data‐driven approaches perform ensemble forecasting using some simple perturbation methods, like the initial condition perturbation and the Monte Carlo dropout. However, their ensemble performance is often limited arguably by the sub‐optimal ways of applying perturbation. We propose a Swin Transformer‐based Variational Recurrent Neural Network (SwinVRNN), which is a stochastic weather forecasting model combining a SwinRNN predictor with a perturbation module. SwinRNN is designed as a Swin Transformer‐based recurrent neural network, which predicts the future states deterministically. Furthermore, to model the stochasticity in the prediction, we design a perturbation module following the Variational Auto‐Encoder paradigm to learn the multivariate Gaussian distributions of a time‐variant stochastic latent variable from the data. Ensemble forecasting can be easily performed by perturbing the model features leveraging the noise sampled from the learned distribution. We also compare four categories of perturbation methods for ensemble forecasting, that is, fixed distribution perturbation, learned distribution perturbation, MC dropout, and multi model ensemble. Comparisons on the WeatherBench data set show that the learned distribution perturbation method using our SwinVRNN model achieves remarkably improved forecasting accuracy and reasonable ensemble spread due to the joint optimization of the two targets. More notably, SwinVRNN surpasses operational IFS on the surface variables of the 2‐m temperature and the 6‐hourly total precipitation at all lead times up to 5 days (Code is available at https://github.com/tpys/wwprediction). Plain Language Summary Ensemble forecasting plays a crucial role in numerical weather prediction (NWP), since a single deterministic model is hard to forecast the chaotic atmosphere conditions. Recent works have begun to explore the data‐driven based ensemble methods due to their rapid prediction speed over traditional NWP. We develop an efficient and effective deep learning model capable of generating large ensemble forecasts with high prediction accuracy and low prediction time cost. The predicted ensemble members have much greater and more reasonable ensemble spread, and better coverage of the ground truth, compared to the prior data‐driven methods. Moreover, our model surpasses the state‐of‐the‐art operational NWP model on the surface atmospheric variables of the 2‐m temperature and the 6‐hourly total precipitation, offering an impressive probability weather prediction baseline. Key Points A transformer‐based variational model called SwinVRNN is developed for medium‐range weather prediction The proposed SwinVRNN can effectively generate large ensemble forecasts with great prediction accuracy and reasonable ensemble spread The model sets a new state‐of‐the‐art among data‐driven models and surpasses the Integrated Forecast System on key atmospheric variables

Weather forecasting is important for science and society. At present, the most accurate forecast system is the numerical weather prediction (NWP) method, which represents atmospheric states as discretized grids and numerically solves partial differential equations that describe the transition between those states 1 . However, this procedure is computationally expensive. Recently, artificial-intelligence-based methods 2 have shown potential in accelerating weather forecasting by orders of magnitude, but the forecast accuracy is still significantly lower than that of NWP methods. Here we introduce an artificial-intelligence-based method for accurate, medium-range global weather forecasting. We show that three-dimensional deep networks equipped with Earth-specific priors are effective at dealing with complex patterns in weather data, and that a hierarchical temporal aggregation strategy reduces accumulation errors in medium-range forecasting. Trained on 39 years of global data, our program, Pangu-Weather, obtains stronger deterministic forecast results on reanalysis data in all tested variables when compared with the world’s best NWP system, the operational integrated forecasting system of the European Centre for Medium-Range Weather Forecasts (ECMWF) 3 . Our method also works well with extreme weather forecasts and ensemble forecasts. When initialized with reanalysis data, the accuracy of tracking tropical cyclones is also higher than that of ECMWF-HRES. Three-dimensional deep neural networks can be trained to forecast global weather patterns, including extreme weather, with accuracy greater than or equal to that of the best numerical weather prediction models.

General circulation models (GCMs) are the foundation of weather and climate prediction 1 , 2 . GCMs are physics-based simulators that combine a numerical solver for large-scale dynamics with tuned representations for small-scale processes such as cloud formation. Recently, machine-learning models trained on reanalysis data have achieved comparable or better skill than GCMs for deterministic weather forecasting 3 , 4 . However, these models have not demonstrated improved ensemble forecasts, or shown sufficient stability for long-term weather and climate simulations. Here we present a GCM that combines a differentiable solver for atmospheric dynamics with machine-learning components and show that it can generate forecasts of deterministic weather, ensemble weather and climate on par with the best machine-learning and physics-based methods. NeuralGCM is competitive with machine-learning models for one- to ten-day forecasts, and with the European Centre for Medium-Range Weather Forecasts ensemble prediction for one- to fifteen-day forecasts. With prescribed sea surface temperature, NeuralGCM can accurately track climate metrics for multiple decades, and climate forecasts with 140-kilometre resolution show emergent phenomena such as realistic frequency and trajectories of tropical cyclones. For both weather and climate, our approach offers orders of magnitude computational savings over conventional GCMs, although our model does not extrapolate to substantially different future climates. Our results show that end-to-end deep learning is compatible with tasks performed by conventional GCMs and can enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system. A hybrid model that combines a differentiable solver for atmospheric dynamics with machine-learning components is capable of weather forecasts and climate simulations on par with the best machine-learning and physics-based methods.

A unique, high-resolution, hydroclimate reanalysis, 40-plus-year (October 1979–September 2021), 4 km (named as CONUS404), has been created using the Weather Research and Forecasting Model by dynamically downscaling of the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis of the global climate dataset (ERA5) over the conterminous United States. The paper describes the approach for generating the dataset, provides an initial evaluation, including biases, and indicates how interested users can access the data. The motivation for creating this National Center for Atmospheric Research (NCAR)–U.S. Geological Survey (USGS) collaborative dataset is to provide research and end-user communities with a high-resolution, self-consistent, long-term, continental-scale hydroclimate dataset appropriate for forcing hydrological models and conducting hydroclimate scientific analyses over the conterminous United States. The data are archived and accessible on the USGS Black Pearl tape system and on the NCAR supercomputer Campaign storage system.

In recent years, deep learning models have rapidly emerged as a stand-alone alternative to physics-based numerical models for medium-range weather forecasting. Several independent research groups claim to have developed deep learning weather forecasts that outperform those from state-of-the-art physics-based models, and operational implementation of data-driven forecasts appears to be drawing near. However, questions remain about the capabilities of deep learning models with respect to providing robust forecasts of extreme weather. This paper provides an overview of recent developments in the field of deep learning weather forecasts and scrutinises the challenges that extreme weather events pose to leading deep learning models. Lastly, it argues for the need to tailor data-driven models to forecast extreme events and proposes a foundational workflow to develop such models.

Many atmosphere and climate phenomena lie in the gray zone between weather and climate: they are not amenable to deterministic forecast, but they still depend on the initial condition. A natural example is medium-range forecasting, which is inherently probabilistic because it lies beyond the deterministic predictability time of the atmosphere, but for which statistically significant prediction can be made, which depends on the current state of the system. Similarly, one may ask the probability of occurrence of an El Niño event several months ahead of time. We introduce a quantity that corresponds precisely to this type of prediction problem: the committor function is the probability that an event takes place within a given time window, as a function of the initial condition. We compute it in the case of a low-dimensional stochastic model for El Niño, the Jin and Timmermann model. In this context, we show that the ability to predict the probability of occurrence of the event of interest may differ strongly depending on the initial state. The main result is the new distinction between probabilistic predictability (when the committor function is smooth and probability can be computed, which does not depend sensitively on the initial condition) and probabilistic unpredictability (when the committor function depends sensitively on the initial condition). We also demonstrate that the Jin and Timmermann model might be the first example of a stochastic differential equation with weak noise for which transition between attractors does not follow the Arrhenius law, which is expected based on large deviation theory and generic hypothesis.

Based on deterministic and probabilistic forecast verification, we investigated the performance of three subseasonal-to-seasonal (S2S) operational models, i.e., the model of the European Centre for Medium-Range Weather Forecasts (ECMWF) and two models of the China Meteorological Administration (CMA1.0 and CMA2.0), in the extended-range forecast of extreme rainfall over southern China (SCER) while considering the modulation of 10–30-day boreal summer intraseasonal oscillation (BSISO2). The Heidke Skill Score (HSS) of the SCER in the ECMWF, CMA2.0, and CMA1.0 models decreased to less than 0.1 at lead times of 13, 9, and 6 days, respectively. Similarly, the useful prediction skill of the BSISO2 index in the ECMWF, CMA1.0, and CMA2.0 models was up to 15, 13, and 8 days in advance, respectively. The BSISO2’s phase error, rather than the amplitude error, determines its prediction skill. The HSS of the BSISO2 index is significantly correlated with that of SCER in all three S2S models, suggesting that the prediction skill of SCER is influenced by that of BSISO2. The ECMWF shows much higher skill than the two CMA models do in predicting the SCER probability changes under the influence of BSISO2 during Phases 5–7, with the useful prediction skill having up to a 10-day lead time. In contrast, CMA1.0 and CMA2.0 can only predict the modulation of BSISO2 on the SCER probability within a week. The prediction skill of BSISO2’s modulation on SCER largely relies on moisture convergence, rather than on moisture advection. This study highlighted the importance of model’s accurate representation of BSISO2 and its associated moisture convergence for improving extended-range forecast of SCER.

Significant advancements in the development of machine learning (ML) models for weather forecasting have produced remarkable results. State-of-the-art ML-based weather forecast models, such as FuXi, have demonstrated superior statistical forecast performance in comparison to the high-resolution forecasts (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF). However, a common limitation of these ML models is their tendency to generate increasingly smooth predictions as forecast lead times increase, which often results in the underestimation of intensities of extreme weather events. To address this challenge, we developed the FuXi-Extreme model, which employs a denoising diffusion probabilistic model (DDPM) to enhance finer-scale details in the surface forecast data generated by the FuXi model in 5-day forecasts. An evaluation of extreme total precipitation (TP), 10-meter wind speed (WS10), and 2-meter temperature (T2M) illustrates the superior performance of FuXi-Extreme over both FuXi and HRES. Moreover, when evaluating tropical cyclone (TC) forecasts based on International Best Track Archive for Climate Stewardship (IBTrACS) dataset, both FuXi and FuXi-Extreme shows superior performance in TC track forecasts compared to HRES, but they show inferior performance in TC intensity forecasts in comparison to HRES.

Root-zone soil moisture (RZSM) is an important variable in land–atmosphere interactions, notably affecting the global climate system. Contrary to satellite-based acquisition of surface soil moisture, RZSM is generally obtained from model-based simulations. In this study, in situ observations from the Naqu and Pali networks that represent different climatic conditions over the Tibetan Plateau (TP) and a triple collocation (TC) method are used to evaluate model-based RZSM products, including Global Land Evaporation Amsterdam Model (GLEAM) (versions 3.5a and 3.5b), Global Land Data Assimilation System (GLDAS) (versions 2.1 and 2.2), and the fifth-generation European Centre for Medium-Range Weather Forecasts reanalysis (ERA5). The evaluation results based on in situ observations indicate that all products tend to overestimate but could generally capture the temporal variation, and ERA5 exhibits the best performance with the highest 𝑅 (0.875) and the lowest unbiased RMSE (ubRMSE; 0.015 m³ m⁻³) against in situ observations in the Naqu network. In the TC analysis, similar results are obtained: ERA5 has the best performance with the highest TC-derived 𝑅 (0.785) over the entire TP, followed by GLEAM v3.5a (0.746) and GLDAS-2.1 (0.682). Meanwhile, GLEAM v3.5a and GLDAS-2.1 outperform GLEAM v3.5b and GLDAS-2.2 over the entire TP, respectively. Besides, possible error causes in evaluating these RZSM products are summarized, and the effectiveness of TC method is also evaluated with two dense networks, finding that TC method is reliable since TC-derived 𝑅 is close to ground-derived 𝑅, with only 6.85%mean relative differences. These results using both in situ observations and TC method may provide a new perspective for the soil moisture product developers to further enhance the accuracy of model-based RZSM over the TP.

Precipitation is the most important driver of the hydrological cycle, but it is challenging to estimate it over large scales from satellites and models. Here, we assessed the performance of six global and quasi-global high-resolution precipitation datasets (European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis version 5 (ERA5), Climate Hazards group Infrared Precipitation with Stations version 2.0 (CHIRPS), Multi-Source Weighted-Ensemble Precipitation version 2.80 (MSWEP), TerraClimate (TERRA), Climate Prediction Centre Unified version 1.0 (CPCU), and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System-Climate Data Record (PERSIANN-CCS-CDR, hereafter PERCCDR) for hydrological modelling globally and quasi-globally. We forced the WBMsed global hydrological model with the precipitation datasets to simulate river discharge from 1983 to 2019 and evaluated the predicted discharge against 1825 hydrological stations worldwide, using a range of statistical methods. The results show large differences in the accuracy of discharge predictions when using different precipitation input datasets. Based on evaluation at annual, monthly, and daily timescales, MSWEP followed by ERA5 demonstrated a higher correlation (CC) and Kling–Gupta efficiency (KGE) than other datasets for more than 50 % of the stations, whilst ERA5 was the second-highest-performing dataset, and it showed the highest error and bias for about 20 % of the stations. PERCCDR is the least-well-performing dataset, with a bias of up to 99 % and a normalised root mean square error of up to 247 %. PERCCDR only show a higher KGE and CC than the other products for less than 10 % of the stations. Even though MSWEP provided the highest performance overall, our analysis reveals high spatial variability, meaning that it is important to consider other datasets in areas where MSWEP showed a lower performance. The results of this study provide guidance on the selection of precipitation datasets for modelling river discharge for a basin, region, or climatic zone as there is no single best precipitation dataset globally. Finally, the large discrepancy in the performance of the datasets in different parts of the world highlights the need to improve global precipitation data products.