A novel approach for assimilating GNSS tropospheric gradient information to improve numerical weather prediction

The intensification and frequency of extreme weather events in the United States present challenges for the traditional numerical weather prediction (NWP) models. New breakthroughs in artificial intelligence (AI) and machine learning (ML) have opened unprecedented opportunities for improving the performance of climate models, particularly the forecast of extreme weather events. In this narrative review, we summarize the latest advances in AI-augmented climate modeling, focusing on the impact of the integration of machine learning on the skill of extreme weather prediction at the U.S. facets of emerging trends, methodological progress and operational implications. We surveyed AI integration with traditional climate models and identified themes from peer-reviewed literature in 2020 to present, including recent advances in neural weather models, hybrid physics-ML approaches, and extreme event detection algorithms. Revolutionary AI models e.g. GraphCast, FourCastNet, and FengWu outperform traditional NWP systems on medium-range forecasting. Deep learning-based methods are particularly promising for extreme event prediction, having achieved record-breaking performance in heatwave, hurricane, and flood detection and prediction using deep convolutional neural networks and transformer-based architectures. Yet, difficulties still exist in the interpretability of models, quantification of uncertainties, and generalization of new extremes. AI-assisted dynamic modulations are game changers in agrometeorological predictions, with superior predictive power for extreme weather events needed for disaster alert systems, risk mitigation and climate change management plans in the USA.

ABSTRACT Accurate photovoltaic (PV) power forecasting is critical for grid integration of PV energy. To address large errors in numerical weather prediction (NWP) data and poor seasonal adaptability of traditional models, this study proposes an intelligent forecasting method integrating Analog Ensemble (AnEn)-based NWP correction and a seasonally segmented convolutional neural network – long short-term memory (CNN-LSTM) multi-model framework. First, AnEn calibrates NWP data to improve meteorological input accuracy; specifically, corrected irradiance reduces root mean square error (RMSE) by 14.62% and mean absolute error (MAE) by 21.56%, while corrected temperature decreases RMSE by 11.09% and MAE by 12.67%. Second, seasonal CNN-LSTM models capture distinct PV generation characteristics across seasons. Experimental results show the proposed method outperforms traditional models: overall RMSE is reduced by 5.34% (from 16.58 to 15.69), MAE by 6.78% (from 11.85 to 11.05), and the coefficient of determination (R2) is improved by 0.47% (from 0.958 to 0.962). This work enhances forecasting accuracy and robustness, with strong practical applicability for power system operation

Abstract Accurately simulating orography‐induced mountain waves over steep terrain, such as the Tibetan Plateau (TP), remains a major challenge for numerical weather prediction (NWP) models due to grid distortions inherent in traditional terrain‐following coordinates. To address this issue, we developed an AI‐driven adaptive mesh refinement (AMR) framework within the Fluidity‐Atmosphere model, which employs a 3D unstructured mesh to mitigate geometric distortions. A Long Short‐Term Memory (LSTM) neural network is integrated to enhance the AMR process, replacing traditional adaptation criteria with data‐driven predictions. A series of idealized 2D and 3D experiments demonstrate that both the traditional AMR and LSTM‐driven approaches reproduce mountain wave dynamics with higher efficiency than fixed mesh. Furthermore, the LSTM model suppresses numerical noise near terrain, preventing spurious over‐refinement. In realistic simulations over the TP, the LSTM‐enhanced model successfully captured the life cycle of mountain waves, reproducing key physical features such as vertical velocity structures, wave amplitude decay, and upstream phase tilt. Comparative tests further revealed efficiency gains of up to 71.4% over fixed meshes and 23.8% over traditional AMR at high resolution, alongside accuracy improvements in vertical velocity, potential temperature, and wave propagation. These findings validate the LSTM‐based AMR framework as a robust and efficient approach for atmospheric simulations over complex terrain. By intelligently allocating computational resources while preserving physical accuracy, this method offers a scalable pathway toward next‐generation atmospheric modeling, with future applications targeted at realistic meteorological conditions over the TP.

Abstract AI‐based climate and weather models provide fast, skillful forecasts yet face a key challenge: predicting future climates while being trained with historical data. We investigate this issue by analyzing boreal winter land temperature biases in AI weather (FourCastNet V2 Small and Pangu Weather) and climate (Ai2 Climate Emulator version 2) models. We evaluate these models during time periods that are significantly more recent than the bulk of their training data, allowing us to assess how well they generalize to more modern conditions. We find that all models produce cold‐biased mean temperatures, resembling climates from 15 to 20 years earlier than their prediction period. Furthermore, FourCastNet's and Pangu's cold bias is strongest for the hottest predicted temperatures, indicating limited training exposure to modern extreme heat events. In contrast, ACE2's bias is more evenly distributed but largest in regions, seasons, and parts of the temperature distribution where historic global warming is most pronounced.

Abstract Polar lows (PLs) are intense mesoscale cyclone systems that rapidly develop and pose risks to coastal infrastructure, shipping, and maritime operations. Hence, making accurate predictions of PL trajectories is crucial. Due to the high dynamics and small scales of PLs (100 km), numerical dynamical models require a high resolution to properly resolve these systems, thereby increasing computational costs. However, such high‐resolution models still face challenges in accurate predictions. To address these issues, this study explores the application of machine‐learning (ML) models, including temporal models and spatiotemporal sequence models for 12‐hour forecasts of PL trajectories. Such algorithms are a considerably cheaper alternative to numerical models. We train the ML models on a high‐resolution reanalysis. The spatiotemporal models trained with key meteorological variables perform better than the temporal models at the first two lead three‐hour time steps. At later steps, spatiotemporal models trained with historical data show growing errors that surpass those of the temporal type. Encouragingly, however, spatiotemporal models trained with historical and future data, especially the combination of lower‐ and upper‐level geopotential height fields, achieve the best predictive accuracy, consistently maintaining the lowest mean distance error (MDE) and outperforming a benchmark error (67.2 km) constituting the error of the reanalysis relative to an expert‐derived PL list (the Noer list). Moreover, introducing a physical constraint loss function associated with the MDE as well as an ensemble method further improves prediction accuracy. These findings demonstrate that ML models can generate fast and accurate PL trajectory forecasts, producing results more quickly than numerical weather prediction models. Incorporating future meteorological variables from numerical models, along with high‐quality trajectory data, further enhances the prediction accuracy of ML models, suggesting potential for improving the operational forecasting of PLs, based on a combination of numerical and ML models.

Abstract. This study investigated the potential of cloud seeding to mitigate extreme rainfall localization (i.e., overseeding) associated with mesoscale convective systems in Japan. Using a numerical weather prediction model, we conducted cloud seeding experiments by artificially increasing ice nuclei concentrations in a double-moment microphysics scheme for the heavy rainfall event in Hiroshima Prefecture, Japan, in August 2014. We examined the sensitivity of rainfall changes to altitude and area of the seeding. The results showed that seeding in the mid–upper troposphere (7.2–8.6 km), where air temperature ranged from −22 to −12 °C, resulted in the most pronounced changes in rainfall amount. At these levels, high supercooled cloud water content and strong updrafts favoured heterogeneous freezing, resulting in a depletion of moisture and suppression of graupel growth. The cloud seeding led to reduced rainfall within the heavy rainfall region and increased rainfall downwind, demonstrating the hypothesized dispersal mechanism of “overseeding”. Expanding the seeding to cover the upstream region of the heavy rainfall area had a greater impact than increasing vertical thickness of the seeding. The most effective seeding configuration (24 km × 24 km area at 7.2 km) achieved an 11.5 % decrease in area-averaged 3-h accumulated rainfall and a maximum reduction of 32 % in 3-h accumulated rainfall over the heavy rainfall region. Future work should consider more realistic representations of seeding substance (i.e., transport, dispersion, and interactions) and explore a wider range of rainfall events to generalize the applicability of this approach.

Abstract Data‐driven models for weather forecasting display impressive computational speed‐up. However, these gains do not include the training phase of the models. This study gathers information from the literature about training and inference to estimate their energy and carbon footprints. Despite being considerably more costly than a physics‐based forecast, the training is rapidly compensated by the savings made during inferences. For a use‐case corresponding to a one‐year usage, data‐driven models are estimated to consume 21 to 1273 times less energy than the physics‐based model. Consequently, for low‐resolution forecasts, data‐driven models bring opportunities to significantly reduce the carbon footprint of weather forecasting.

The nowcasting of extreme rainfall poses significant daily challenges on a global scale, especially in vulnerable regions of the Global South. Conventional Numerical Weather Prediction models often fail to deliver accurate and timely forecasts for extreme weather events, exacerbating socioeconomic inequalities and increasing climate vulnerability. Deep learning approaches present a promising opportunity to uncover more precise predictive patterns; however, their application remains constrained by the high computational costs associated to their large parameter spaces. This study evaluates the effectiveness of the MS-RNN framework for improving computational efficiency and predictive accuracy in extreme precipitation nowcasting, using real weather radar data from the TAASRAD19 and Rio de Janeiro datasets. While the framework has been extensively validated both theoretically and experimentally in other scenarios, this work examines its application to real radar data. Metrics related to sustainability, such as energy consumption, [Formula: see text] emissions, and water usage, have not been calculated in this specific context and are rarely addressed in current literature. Our findings demonstrate the potential of the solution to enhance computational efficiency maintaining predictive performance when applied to real weather radar data, supporting sustainable and accessible AI solutions for climate resilience in resource-limited regions.

Fine-scale near-surface wind field prediction is essential for a wide range of applications. However, most operational and AI-based weather models operate at kilometer-scale resolution, where terrain-induced wind features such as slope jets, flow deflection, and recirculation are systematically averaged out. Here we introduce FuXi-CFD, a machine learning-based framework designed to generate detailed three-dimensional (3D) near-surface wind fields at 30-meter horizontal resolution, using only coarse-resolution atmospheric inputs and high-resolution terrain information. The model is trained on a large-scale dataset generated via computational fluid dynamics (CFD), encompassing a wide range of terrain types and inflow conditions. Although relying only on horizontal wind inputs, FuXi-CFD infers the full 3D wind fields-including latent variables such as vertical velocity and turbulence-related features. It achieves CFD-comparable accuracy while reducing inference time from hours to seconds. Notably, the model also generalizes well to real-world conditions, as demonstrated by consistent performance against independent wind-tower observations. This capability enables real-time wind field reconstruction for terrain-sensitive applications such as wind turbine siting, power forecasting, and wildfire spread modeling.

Accurate runoff forecasting is essential for effective water resource management, hydropower operation, and flood risk mitigation. In this study, daily inflow runoff in the Xin’an River Basin, eastern China, was simulated using four ensemble learning models: Gradient Boosting Decision Tree (GBDT), XGBoost, CatBoost, and Stacking. Among them, the CatBoost model achieved the best performance, with a correlation coefficient (CC) exceeding 0.97, Nash–Sutcliffe efficiency (NSE) above 0.95, and reduced RMSE and MAE compared with the currently operational hydrological model. To extend the forecast lead times, two hydro–meteorological coupled models were developed by integrating the CatBoost model with a single numerical weather prediction model (EC) and a dynamically weighted multi-model ensemble precipitation forecast system (OCF). The coupled models were evaluated for lead times up to 240 h. The forecast skill value was highest within 96 h, with CC values above 0.80 and NSE around 0.50. The OCF-coupled model demonstrated improved reliability for lead times of 48–96 h, whereas the EC-driven forecasts performed better within the first 48 h. Case studies during the 2021–2022 flood seasons confirmed that the coupled framework accurately reproduced flood evolution and peak discharge dynamics, demonstrating its practical value for medium-range runoff forecasting in humid river basins.

Persistent knowledge gaps in precipitation microphysics, particularly the nonlinear coupling between microphysical process hierarchies and raindrop size distribution (DSD) variability, keep introducing systemic uncertainties into precipitation retrievals and model simulations. Here, we address this challenge through a unified framework that integrates observations from China's national-scale disdrometer network (1,031 sites) and 10-year global dual-frequency precipitation satellite dataset. First, a region-independent DSD continuum characterized by a universal linear relationship between raindrop diameter and concentration across diverse climatic zones is identified, extending and refining the conventional maritime-like and continental-like category. Then, we quantify the vertical stratification of microphysical processes in shaping and shifting this continuum. Implementations of our findings to reduce biases in current microphysics parameterizations are proposed and discussed. This study advances our fundamental understanding of the apparent heterogeneity yet inherent homogeneity in the microphysics of heavy precipitation, providing mechanistic insights to improve the performance of weather and climate models.

Abstract The development of deep convection and the timing of storm convection initiation over land are generally poorly represented by weather and climate models, probably due to the poor representation of the interaction between shallow and deep convection. The present work aims to present a prognostic model for the shallow‐to‐deep transition of Amazonian convection in which the shallow and deep clouds are described under a unified framework. Three types of clouds are considered in the present model: shallow cumuli, cumulus congestus, and deep cumulonimbus clouds. The model is based on the idea that the shallow‐to‐deep transition of atmospheric convection is primarily controlled by the interaction between updrafts and passive cloud volumes (non‐convective cumulus cloud volumes in the decaying stage) from which moist air can be entrained in the updrafts. In this framework, the cold pools accelerate the shallow‐to‐deep transition due to a larger cloud‐updraft interaction, allowing the updrafts to be less susceptible to entrainment of dry environmental air. A dynamical model for the diurnal cycle of shallow and deep is obtained and tested against idealized large‐eddy simulations. For idealized cases of Amazonian atmospheric convection, it is shown that the model represents the transition from shallow to deep convection reasonably well, showing that the model could lead to improvements in the forecast for storm convection initiation.

Climate foundation models (FMs) have achieved remarkable success in weather forecasting, yet exhibit puzzling performance gaps between tasks: deterministic field prediction rivals operational numerical weather prediction, while extreme event detection lags significantly behind. We provide a theoretical explanation through the lens of in-context learning (ICL). Extending the ICL characterization framework to spatiotemporal climate data, we prove the Climate Prediction Dichotomy Theorem: every natural climate task falls into exactly one of two complexity categories. Type A (ICL-Easy) tasks—including temperature, pressure, and wind field prediction—admit additive sufficient statistics enabling attention-based computation with sample complexity nICL = Θ(nERM). Type C (ICL-Hard) tasks—including extreme event detection, tipping point identification, and compound event localization—require combinatorial sufficient statistics that provably exceed the computational capacity of constant-depth polynomial-size transformers when the number of simultaneous events J exceeds a threshold J* = O(log log n) ≈ 3–5. We establish the predictability horizon constraint: weather forecasting is ICL-Easy for lead times τ < τL ≈ 14 days, while climate statistics remain ICL-accessible but individual trajectories are fundamentally unpredictable. Our analysis yields six testable predictions about climate FM behavior and five deployment guidelines distinguishing when ICL suffices versus when fine-tuning is required. The dichotomy provides a principled foundation for understanding why Pangu-Weather and GraphCast excel at field prediction while struggling with event detection—and guides the design of next-generation climate AI systems.

Abstract. The combined use of independent soil moisture data from radar and radiometer measurements in data assimilation (DA) systems is expected to yield synergistic performance gains due to their complementary strengths. This study evaluates the impact of simultaneously assimilating soil moisture retrievals from ASCAT (Advanced SCATterometer) and SMAP (Soil Moisture Active Passive) into the Korean Integrated Model (KIM) using a weakly coupled DA framework based on the National Aeronautics and Space Administration's Land Information System (LIS). The Noah land surface model (LSM) within LIS, which is the same as that used in KIM, is used to simulate land surface states and assimilate soil moisture retrievals. The impact of soil moisture DA is evaluated using independent reference datasets, assessing its influence on soil moisture analysis and numerical weather prediction performance. Overall, assimilating single-sensor soil moisture data, ASCAT or SMAP, into the LSM improves global soil moisture analysis accuracy by 4.0 % and 10.5 %, respectively, compared to the control case without soil moisture DA, achieving the most significant enhancements in croplands. Relative to single-sensor soil moisture DA, multi-sensor soil moisture DA yields more balanced skill enhancements for both specific humidity and air temperature analyses and forecasts. The most pronounced synergistic improvements by simultaneously assimilating both soil moisture products are observed in the 2 m air temperature analysis and forecast, especially when both soil moisture products have a positive impact. Precipitation forecast skill also improves with multi-sensor soil moisture DA, although the improvements are not consistent across regions and events. This paper discusses remaining issues for future studies to further improve the weather prediction performance of the KIM-LIS multi-sensor soil moisture DA system.

Dust sharply degrades visibility in arid and semi-arid regions, yet operational forecasting remains challenged by near-surface process errors in numerical weather prediction (NWP) and the poor generalization of purely data-driven models. We present a physics-guided machine learning (PGML) framework that post-processes European Centre for Medium-Range Weather Forecasts (ECMWF) forecasts to predict five ordinal visibility grades over the Kumtag Desert. A dust-lifecycle feature library (emission, vertical mixing, transport, wet scavenging) is coupled with an ordinal LightGBM architecture. On an independent test period, the model attains quadratic weighted kappa (QWK) of 0.26 (0-24h), 0.17 (24-48h), and 0.18 (48-72h), with mean absolute error (MAE) of 0.48-0.56; gains versus data-only baselines increase with forecast horizon. Ablation experiments show that physics priors can effectively improve visibility prediction accuracy-reducing MAE by up to 10% and sustaining QWK beyond 24h by constraining non-physical drift. Accordingly, the PGML visibility predictions show improved performance relative to data-only baselines. SHAP analysis reveals a forecast-horizon-dependent mechanistic shift from emission/surface-layer dynamics to stability-controlled vertical mixing, consistent with dust dynamics. The framework offers an interpretable, transferable paradigm for physics-constrained environmental forecasting.

Abstract. Snow cover is a key variable for climate monitoring and hydrological applications, yet optical satellite observations are strongly limited by cloud contamination, particularly during winter. The EUMETSAT H SAF H34 snow product derived from MSG-SEVIRI uses the unique 15-minute temporal resolution over the full SEVIRI disc to clear the clouds, but persistent clouds still cause substantial data gaps. In this study, we present a cloud-gap reconstruction framework that combines Numerical Weather Prediction data with machine learning to infer snow presence beneath cloud-covered pixels in the H34 product. Skin temperature, snow depth, and snow temperature fields from the Integrated Forecast System (IFS) were used as physically consistent predictors and resampled to the H34 grid, together with elevation information from SRTM. An XGBoost-based model was trained using cloud-free H34 snow observations and applied exclusively to cloud-contaminated pixels to estimate the probability of underlying snow presence. Pixels exceeding an 80% probability threshold were reclassified as snow. The approach was applied to the winter seasons of 2024 and 2025 and validated over the European Alps using in-situ snow observations from World Meteorological Organization (WMO) stations. Evaluation using probability of detection (POD), false alarm ratio (FAR), and overall accuracy (ACC) shows a clear improvement in snow detection under cloudy conditions, with a significant reduction in missing observations. Compared to conventional temporal gap-filling methods, the proposed framework reduces reliance on temporal interpolation by directly exploiting physically meaningful meteorological information, while preserving the high temporal resolution advantage of MSG-SEVIRI.

Abstract. Extratropical windstorms pose a major hazard in Northern Europe, with damage primarily arising from the combined effects of high sustained near-surface winds and extreme gusts. While future changes in mid-latitude storms are expected under climate warming, the implications for wind-related impacts remain uncertain. In this study, we investigate the response of thermodynamic warming to an intense historical storm using Storm Anatol, which severely affected Denmark on 3 December 1999, as a representative case. The storm is simulated using the convection-permitting numerical weather prediction model HARMONIE-AROME within a pseudo-global warming framework. Uniform temperature perturbations are applied throughout the atmosphere, sea surface, and skin layers, while specific humidity is adjusted to maintain relative humidity. Changes in wind speed, gusts, and the spatial and temporal extent of damaging wind conditions are analysed across a range of warming scenarios. To quantify integrated wind exposure, we employ a new cumulative metric applicable to both wind speed and gust diagnostics, referred to as the Cumulative Wind Exposure Index. The simulations show a systematic intensification of near-surface wind and gust speeds with increasing temperature, accompanied by an expansion in the spatial footprint and duration of extreme wind conditions. The cumulative wind exposure increases markedly in the warmer scenarios relative to the historical simulation. When interpreted in the context of established wind–damage relationships, these changes imply substantially enhanced potential for wind-related impacts. Overall, the results demonstrate that thermodynamic warming alone can significantly amplify windstorm exposure, highlighting the importance of considering compound wind characteristics, when assessing future wind hazards and their impacts in Northern Europe.

The hyperspectral infrared observations of the Geostationary Interferometric Infrared Sounder (GIIRS) on the Fengyun-4A (FY-4A) satellite are an important data source for numerical weather prediction (NWP) assimilation. However, there are systematic differences between observed and simulated brightness temperatures (i.e., the observation increments contain predictable systematic bias components). To address the issue that traditional linear methods struggle to capture the nonlinear relationships between biases and forecast predictors, this study proposes an intelligent bias correction method that integrates ensemble learning and explainable artificial intelligence. First, the entropy reduction method is used to select 69 mid-wave channels. Then, Random Forest, XGBoost, LightGBM, Decision Tree, and Extra Tree are used as base learners to construct a weighted average ensemble model. Training and validation are conducted using high-frequency clear-sky observation data from FY-4A/GIIRS during Typhoon Lekima. The results show that: (1) the ensemble learning correction method outperforms single models and traditional offline methods, with root mean square errors of brightness temperature bias of less than 0.9209 K for the training set and 1.4447 K for the test set; (2) Shapley Additive Explanations (SHAP)-based interpretability analysis reveals the contribution and nonlinear influence mechanisms of factors such as longitude, atmospheric thickness, surface temperature, and total precipitable water on bias correction. This study provides an intelligent bias correction framework with both high precision and explainability, offering a reference for the bias correction and assimilation applications of hyperspectral satellite observations like GIIRS.

Enhancing rainfall prediction accuracy through image fusion of radar and numerical weather prediction models