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 predict global weather patterns days advance. First studies show promise but the lack of a common dataset and evaluation metrics make inter-comparison between difficult. Here we present benchmark medium-range forecasting, topic high scientific interest atmospheric computer scientists alike. We provide...

Abstract It is shown that it possible to emulate the dynamics of a simple general circulation model with deep neural network. After being trained on model, network can predict complete state several time steps ahead—which conceptually making weather forecasts in world. Additionally, after initialized an arbitrary state, through repeatedly feeding back its predictions into inputs create climate run, which has similar statistics model. This run shows no long‐term drift, even though...

Weather forecasts are inherently uncertain. Therefore, for many applications only considered valuable if an uncertainty estimate can be assigned to them. Currently, the best method provide a confidence individual is produce ensemble of numerical weather simulations, which computationally very expensive. Here, we assess whether machine learning techniques alternative approach predict forecast given large‐scale atmospheric state at initialization. We propose based on deep with artificial...

Abstract. Recently, there has been growing interest in the possibility of using neural networks for both weather forecasting and generation climate datasets. We use a bottom–up approach assessing whether it should, principle, be possible to do this. relatively simple general circulation models (GCMs) PUMA PLASIM as simplified reality on which we train deep networks, then predicting model at lead times few days. specifically assess how complexity affects network's forecast skill dependent is...

Abstract Ensemble weather forecasts enable a measure of uncertainty to be attached each forecast, by computing the ensemble's spread. However, generating an ensemble with good spread‐error relationship is far from trivial, and wide range approaches achieve this have been explored—chiefly in context numerical prediction models. Here, we aim transform deterministic neural network forecasting system into system. We test four methods generate ensemble: random initial perturbations, retraining...

Abstract. Neural networks are able to approximate chaotic dynamical systems when provided with training data that cover all relevant regions of the system's phase space. However, many practical applications diverge from this idealized scenario. Here, we investigate ability feed-forward neural (1) learn behavior incomplete and (2) influence an external forcing on dynamics. Climate science is a real-world example where these questions may be relevant: it concerned non-stationary system subject...

Abstract Global warming projections point to a wide range of impacts on the climate system, including changes in storm track activity and more frequent intense extreme weather events. Little is however known whether how global may affect atmosphere's predictability thus our ability produce accurate forecasts. Here, we combine state‐of‐the‐art ensemble prediction model show that, business‐as‐usual 21st century setting, could significantly change atmosphere, defined here via expected error...

Abstract. Skillful forecasts of extreme weather events have a major socioeconomic relevance. Here, we compare two complementary approaches to diagnose the predictability weather: recent developments in dynamical systems theory and numerical ensemble forecasts. The former allows us define atmospheric configurations terms their persistence local dimension, which provides information on how atmosphere evolves from given state interest. These metrics may be used as proxies for intrinsic...

Abstract The accurate prediction of extreme weather events is an important and challenging task, has typically relied on numerical simulations the atmosphere. Here, we combine insights from forecasts with recent developments in dynamical systems theory, which describe atmospheric states terms their persistence ( θ −1 ) local dimension d ), inform how atmosphere evolves to a given state interest. These metrics are intuitively linked intrinsic predictability atmosphere: highly persistent,...

Abstract. Atmospheric large-scale patterns strongly determine Greenland’s regional climate through air mass advection and local weather conditions, making them essential to understand atmospheric variability. This study analyses the occurrence of during two distinct warming periods recent past that we identify objectively in climatological data. The first period lasted from 1922 1932 an average temperature increase 2.9 °C across all stations considered for this study. second 1993 2007 had...

Skilful weather forecasts help users make sound decisions when faced with potentially hazardous climatic conditions. However, this beneficial result may be reduced or negated in the absence of an effective communication forecast uncertainty. On average, skill improves for shorter lead times, which implies that we expect differences between successive forecasts. While there is a vast literature on and visualisation uncertainty, little attention has been dedicated to communicating changes...

Altitude-driven gradients of air temperature, humidity, wind, and surface mass balance play a critical role in understanding glacier-climate interactions, particularly regions rapid environmental change like the Arctic. In this study, we compare datasets from Alfred Wegener’s last expedition to west coast Greenland 1930/31 with modern measurement network established at same locations 2022. This unique comparison offers insights into how atmospheric glacial conditions have changed...

The tropopause is a sensitive indicator of both radiative and dynamic changes in the atmospheric climate system. This study presents an analysis lapse-rate (LRT) trends using remote-sensing satellite data for period 2002-2023. evaluation performed GNSS radio occultation measurements, which are particularly well suited observing temperature region with high vertical resolution global coverage. In addition, provides long-term stable measurements that allow robust detection trends. Our results...

A probabilistic machine learning method is applied to icing related production loss forecasts for wind energy in cold climates. The employed method, called quantile regression forests, based on the random forest algorithm. Based performed tests data from four Swedish parks available two winter seasons, it has been shown produce valuable forecasts. Even with limited amount of training and test that were used study, estimated forecast uncertainty adds more value when compared a deterministic...

Ice-growth on wind-turbines can lead to a large reduction of energy production. Since ice-growth the turbines is not part standard weather prediction data, forecasts power production have errors when occurs. We propose statistical method based random-forest regression predict loss induced by ice-growth. It takes as input both regional and on-site measurements, predicts relative up 42 hours ahead in order improve for next-day The trained past significantly outperforms simple - but also useful...

Abstract. Recently, there has been growing interest in the possibility of using neural networks for both weather forecasting and generation climate datasets. We use a bottom-up approach assessing whether it should, principle, be possible to do this. relatively simple General Circulation Models (GCMs) PUMA PLASIM as simplified reality on which we train deep networks, then predicting model at lead times few days. specifically assess how complexity affects network's forecast skill, dependent...

The use of data-driven decision support by public agencies is becoming more widespread and already influences the allocation resources. This raises ethical concerns, as it has adversely affected minorities historically discriminated groups. In this paper, we an approach that combines statistics approaches with dynamical modeling to assess long-term fairness effects labor market interventions. Specifically, develop a model investigate impact decisions caused employment authority selectively...

Abstract Modeled wintertime precipitation over the Atlantic Gulf Stream region is shown to be sensitive horizontal resolution of driving Global Circulation Model (GCM). By contrasting simulations with EC‐Earth GCM a range resolutions (T159, T319, T799), it that especially extremes become more populated if higher. Higher also appears strengthen communication from sea surface toward troposphere. With increasing resolution, deep convection region, diagnosed via wind‐convergence and vertical...

Abstract. Creating spatially coherent rainfall patterns with high temporal resolution from data lower is necessary in many geoscientific applications. From a statistical perspective, this presents high- dimensional, highly underdetermined problem. Recent advances machine learning provide methods for such probability distributions. We test the usage of generative adversarial networks (GANs) estimating full distribution spatial resolution, conditioned on field resolution. The GAN trained radar...

Abstract To extract the most information from an ensemble forecast, users would need to consider possible impacts of every member in ensemble. However, not all have resources do this. Many may opt only mean and possibly some measure spread around mean. This provides little about potential worst-case scenarios. We explore different methods scenarios for a given definition severity impact: taking worst ensemble, calculating N members, two that use statistical tool known as directional...

Abstract Recently, there has been a surge of research on data-driven weather forecasting systems, especially applications based convolutional neural networks (CNNs). These are usually trained atmospheric data represented regular latitude–longitude grids, neglecting the curvature Earth. We assess benefit replacing standard convolution operations with an adapted operation that takes into account geometry underlying (SphereNet convolution), specifically near poles. Additionally, we effect...

Abstract. Neural networks are able to approximate chaotic dynamical systems when provided with training data that covers all relevant regions of the system's phase space. However, many practical applications diverge from this idealised scenario. Here, we investigate ability neural to: 1) learn behaviour incomplete data, and 2) influence an external forcing on dynamics. Our analysis is performed Lorenz63 Lorenz95 models. We show trained covering only part space struggle make skillful...