Atmospheric and Oceanic Sciences

Abstract
Decision-makers rely on weather forecasts to plant crops, manage wildfires, allocate water and energy and prepare for weather extremes1–3. Today, such forecasts enjoy unprecedented accuracy as far as 2 weeks thanks to steady advances in physics-based dynamical models and data-driven artificial intelligence models4–10. However, model skill drops precipitously at subseasonal timescales (2–6 weeks ahead) due to compounding errors, systemic model biases and the chaotic nature of the atmosphere11–15. To counter this degradation, we introduce probabilistic bias correction, a machine learning framework that substantially reduces systematic error by learning to correct historical probabilistic forecasts. When applied to the leading dynamical and artificial intelligence models from the European Centre for Medium-Range Weather Forecasts (ECMWF), probabilistic bias correction doubles the modest subseasonal skill of the AI Forecasting System16 and improves the skill of the operationally debiased dynamical model17 for 91% of pressure, 92% of temperature and 98% of precipitation targets. We designed probabilistic bias correction for operational deployment, and, in the 2025 real-time forecasting competition of ECMWF18, its global forecasts were the best for all weather variables and lead times, outperforming the dynamical models from six operational forecasting centres, an international dynamical multimodel ensemble, the AI Forecasting System of ECMWF and the forecasting systems of 34 teams worldwide. These probabilistic skill gains translate into more accurate prediction of extreme events and have the potential to improve agricultural planning, energy management and disaster preparedness in vulnerable communities19. The AI framework probabilistic bias correction is shown to correct weather forecast errors 2 to 6 weeks ahead and boost forecast skill, making it promising for improved extreme-weather prediction.

Abstract
Abstract Arctic sea ice decline has been observed across all seasons and is projected to continue in the future. Relative to historical observations, climate models show their largest biases and uncertainties during winter, when ice cover strongly regulates air–sea heat exchange, ocean mixing and atmospheric circulation. Palaeoclimate records offer opportunities to evaluate model performance beyond the historical period, but applications have remained largely qualitative. Here we compile an Arctic-wide database of mid-Holocene (5–7 thousand years ago) biomarkers to assess sea ice simulations of this past warm period. Using a recently developed proxy system model, we present a Bayesian palaeoclimate model weighting scheme that refines the multi-model mean winter sea ice extent through fully quantitative proxy-model comparisons, accounting for model interdependence and non-stationary proxy seasonality. We demonstrate that, by adjusting the contribution of each model to the ensemble mean based on proxy evidence, model weights can eliminate biases and reduce uncertainties by 85%, to the level of interannual variability. Weighted projections using the same climate models indicate that future winter sea ice extent may be nearly 1 million km 2 lower than prior estimates across forthcoming decades and possible emissions scenarios, with broad implications for Arctic ecosystems, mid-latitude weather, and cryosphere stability.

Abstract
Abstract Accurate estimation of fine particulate matter (PM2.5) and characterization of precursor-response relationships are crucial for human health protection and air quality management. However, current approaches remain limited by the high computational cost and uncertainties of chemical transport models (CTMs), as well as the sample imbalance and limited interpretability of observation-based machine learning methods. Here, we developed Fusion-DeepRSM, an observation-constrained extension of the Deep Learning Response Surface Model (DeepRSM) framework that integrates CTM simulations, multisource satellite observations, and ground-based measurements to derive spatially continuous PM2.5 concentrations and characterize precursor-PM2.5 response relationships. Results from ground-based observations and CTM-based testbed experiments show that Fusion-DeepRSM effectively alleviates the sample imbalance in conventional models and mitigates PM2.5 overestimation in rural regions, with an average reduction of 4.13 μg/m3 while maintaining high predictive accuracy (R2 = 0.872 and RMSE = 11.82 μg/m3). In addition, Fusion-DeepRSM successfully quantifies precursor-PM2.5 response relationships, revealing pronounced seasonal and regional heterogeneity consistent with established atmospheric chemistry understanding. With high computational efficiency, Fusion-DeepRSM further quantifies the effects of meteorological conditions and pollutant concentrations on PM2.5 variability. During the nationwide PM2.5 decline from 2020 to 2023, unfavorable meteorology partially offset pollutant-related improvements (+1.2 versus −1.8 μg/m3), while changes in precursor pollutant conditions played an important role in the February 2023 PM2.5 rebound in the Beijing–Tianjin–Hebei region. These findings demonstrate the strong potential of Fusion-DeepRSM for air quality modeling and region-specific pollution mitigation.

Abstract
Abstract This study analyzes a novel nine-member ensemble of convection-permitting regional climate simulations for Germany to assess projected changes in temperature and precipitation extremes under two future global warming levels, +2 and +3 K. The robustness of the climate change signals within the ensemble is also evaluated to ensure the reliability of the results. Throughout the twenty-first century, temperature-related extremes are projected to intensify, with nighttime extremes showing larger increases than daytime extremes. This amplification of nocturnal warming implies elevated risks to human health and ecosystems. While annual precipitation sums in Germany are projected to change only slightly and are mostly not statistically significant, the frequency and intensity of intense and very extreme precipitation events increase substantially. This change signal is particularly evident in summer, despite a predominantly decreasing seasonal mean rainfall. Such contrasting changes indicate a future characterized by generally drier summers interrupted by heavy rainfall events, with implications for both droughts occurrence and flood risk estimates. Temperature-related extremes show high robustness across almost all of Germany, whereas precipitation-related extremes exhibit greater variability, reflecting stronger natural variability and model uncertainty. The high-resolution simulations reveal added value for temperature-based indices, including more detailed spatial patterns of time averages and changes in specific areas, together with reduced uncertainty as indicated by ensemble spread and signal-to-noise ratios. These new climate model projections offer a robust scientific basis and a high potential for developing adaptation strategies tailored to small-scale landscape units.

Abstract
Forcing together with radiative responses and natural variability has sharply increased Earth’s energy imbalance (EEI), a key driver of climate change, between 2003 and 2023. Declining anthropogenic aerosol emissions and increasing greenhouse gas concentrations are the leading contributors to the forcing. However, observation-based and model-based forcing estimates disagree on the magnitude of aerosol forcing, particularly for effective radiative forcing from aerosol-cloud interactions (ERF ACI ). Here, we show that two independent observation-based estimates of forcing trends achieve unexpectedly good closure between 2003 and 2023. The corresponding model-based trend is within the uncertainty range of observations but substantially smaller, with most of the difference due to models’ underestimation of cloud fraction adjustment to aerosols. Our results suggest a prominent role of aerosols in the recent EEI increase and stronger total historical ERF ACI and higher equilibrium climate sensitivity than the current consensus, with a likely range between 2.8° and 4.8°C.

Abstract
Abstract Nitrous oxide (N2O) is the third-largest contributor to anthropogenic greenhouse gas forcing and the dominant anthropogenic ozone-depleting emission. While mitigation can flatten emissions, there is no clear pathway to eliminate residual emissions from agriculture and other diffuse sources, a key challenge for climate stabilization. The potential role of atmospheric removal has not been holistically assessed, yet N2O’s high potency and long atmospheric lifetime indicate the potential for a durable climate benefit. This review synthesizes emerging approaches for atmospheric N2O removal, organized by reaction mechanism (thermal destruction and thermocatalysis, photolysis, gas-phase radical reactions, photocatalysis, and biological reduction) and deployment format (closed versus open system). Each is evaluated for technical efficacy, climate viability, environmental and social viability, and economic plausibility using lower-bound cost models. No approach yet satisfies all four: gas-phase radical reactors are precluded by their kinetics, thermocatalytic and photocatalytic reactors require ten- to hundred-fold performance gains, and open-system approaches have the lowest cost floors but face the greatest measurement and governance challenges. These results identify the performance thresholds and broader constraints that govern feasibility across pathways, providing a foundation for a research agenda and for assessing the role of N2O removal alongside emissions mitigation.

Abstract
Abstract Drought early warning increasingly relies on high-resolution monitoring, but the skill of daily soil moisture forecasts at lead times of 1–2 weeks driven by real-time meteorological ensembles has rarely been evaluated. The Hydrological Sub-seasonal Forecasting System (HS2S) extends an established national drought-monitoring framework into high-resolution probabilistic soil moisture forecasts up to two weeks ahead for Germany at 0.015625° resolution. We evaluated a daily-initialized HS2S configuration using the 51-member real-time forecast ensemble from the European Centre for Medium-Range Weather Forecasts (ECMWF) during the exceptionally dry spring and summer of 2025 against the operational, model-based German Drought Monitor (GDM), generated within the same mHM framework using observation-based meteorological forcing. The assessment combined drought onset/recovery transition diagnostics with probabilistic (Brier skill score) and categorical verification against an initialized two-state Markov persistence benchmark, and continuous verification against an mHM soil-moisture climatology. HS2S reproduced the timing and mean magnitude of GDM-indicated drought onset and recovery. It showed positive probabilistic skill relative to the initialized Markov benchmark and positive continuous skill relative to the mHM climatology at both lead times, while improving the critical success index for moderate drought (SMI < 0.20) by 15.0 and 15.5 percentage points at lead weeks 1 and 2, respectively. Given HS2S and the GDM share the same mHM–SMI framework and initial states, the verification is model-consistent and evaluates skill in anticipating the GDM-indicated drought state rather than validating absolute soil-moisture accuracy. The added value of real-time meteorological forecasts relative to historical meteorological sequences was not assessed. Within this operationally relevant framework, HS2S showed forecast skill up to two weeks ahead during the 2025 drought, highlighting the potential of high-resolution probabilistic soil-moisture forecasts to complement national drought monitoring with forward-looking information.

Abstract
Abstract The zenith total delay (ZTD) retrieved from 5G communication base stations represents a novel observation source for numerical weather prediction (NWP), leveraging the extensive global deployment of 5G infrastructure. In this study, the direct assimilation of 5G‐ZTD observations into a high‐resolution NWP system was accomplished for the first time. Two precipitation events were selected to evaluate the impact of 5G‐ZTD assimilation on short‐term precipitation forecasting. Assimilating 5G‐ZTD data corrected the dry bias in the initial water vapor field and improved precipitation forecasts in both cases, demonstrating its practical utility for short‐term heavy rainfall forecasts and operational potential given its low cost and extensive coverage.

Abstract
Extreme precipitation poses severe societal and economic risks under elevated atmospheric carbon dioxide (CO2), yet current understanding of how it will respond to various future CO2 removal scenarios remains limited. Focusing on East Asia, a region highly vulnerable to monsoonal dynamics and anthropogenic influences, here we assess the hysteresis of extreme precipitation responses under four CO2 concentration thresholds and two emission reduction rates using climate model simulations. The extreme precipitation indices based on intensity and frequency exhibit an increase during the CO2 peak, but fail to fully return to their initial states even after CO2 levels decline to their initial levels. This hysteresis behavior is amplified by delayed and slower emission reductions. The hysteresis is linked to increases in mean monsoon precipitation associated with hysteresis in tropical Pacific warming patterns and subsequent atmospheric teleconnections. Moisture budget analysis further indicates that enhanced precipitation is attributed to increased moisture transport into East Asia, driven by early-summer rises in specific humidity (thermodynamic factor) through mean circulation and late-summer intensification of monsoon circulation (dynamic factor). These findings reveal the mechanisms shaping East Asian extreme precipitation under different carbon removal pathways and highlight the critical role of both timing and pace of emission reductions in determining regional climate outcomes. Across carbon removal experiments, East Asian extreme precipitation can remain elevated as CO2 declines, with later and slower CO2 removal producing stronger hysteresis associated with delayed tropical Pacific warming and changes in the monsoon circulation.

Abstract
Abstract. Deterministic model predictions can struggle to adequately capture extreme events such as floods and droughts, which are of particular relevance in hydrology. This limitation arises because deterministic models collapse the conditional runoff distribution to a single point estimate. Probabilistic modeling provides a way to address this issue by explicitly representing uncertainty and assigning non-zero probabilities to a range of possible outcomes, including rare and extreme events, thereby capturing the full range of plausible hydrological responses. Motivated by this perspective, we examine whether probabilistic Long Short-Term Memory (LSTM) models improve the representation of extreme events in rainfall–-runoff simulations across Switzerland. Overall, the probabilistic models show good calibration, although some miscalibration remains for the extremes. Differences between models mainly manifest in how uncertainty is distributed: some approaches produce narrower and lighter-tailed distributions, while others yield broader distributions with heavier tails. These trade-offs highlight that probabilistic models differ not only in sharpness but also in how their calibration for rare events. We observe this tradeoff also in models' accuracy metrics. When evaluating the mean of the probabilistic predictions using the Nash–Sutcliffe efficiency (NSE), none of the probabilistic approaches outperform the deterministic LSTM in terms of average predictive accuracy. However, a clear advantage over the determinsitc models emerges when focusing on the tail of the discharge distribution. For the most extreme events (top 0.1 % of the discharge distribution), the deterministic LSTM underestimates more than 90 % of observed values (since it provides estimates of an expectation), whereas probabilistic predictions can capture a substantially larger fraction (67 %) of these extremes within their upper predictive bounds. Building on the additional information provided by probabilistic runoff predictions, we further show how they can be translated into actionable flood warnings using reinforcement learning. To this end, we introduce a Flood Risk Communication Agent (FRiCA) that operates on probabilistic runoff predictions and learns decision rules for issuing warnings of varying intensity. The FRiCA is implemented as an LSTM-based policy network and is trained by rewarding correct warning levels while penalizing the underestimation of flood severity. Results indicate that the FRiCA outperforms simple fixed heuristics, such as issuing warnings based on the predictive mean or a fixed high quantile (e.g., the 99th percentile). While this behavior already demonstrates the potential of reinforcement learning for improved flood risk communication, it also motivates further exploration of better reward design and policy network definition for context-dependent decision policies that adapt to varying hydrological and societal contexts.