Results show DySCo achieves superior dynamical consistency with the coarse GCM trajectories, essentially applying a minimal, causal correction to the GCM, preserving top statistical performance comparable to state-of-the-art unsupervised models.
Abstract
Regional climate risk assessment is critical for applications such as infrastructure design, disaster forecasting, and insurance resource allocation. However, estimating regional (i.e., high-spatial-resolution) risk with global climate models (GCMs) remains computationally prohibitive, which has driven the development of downscaling methods for coarse GCM outputs. Downscaling is vital for rare events, since quantifying their extreme properties requires high spatial resolution and very long GCM simulations. These methods non-intrusively increase GCM resolution while correcting statistical biases from unresolved fine-scale processes, thereby improving the accuracy of extreme event statistics with long return periods. A key challenge is preserving dynamical consistency, as freely evolving GCM trajectories are not expected to track the observational dataset used for training the correction operator. This is critical for causal extreme event analyses, where storyline-based risk assessment, i.e., extreme event catalogs, is necessary for effective planning. We address this challenge by introducing Dynamically and Statistically Consistent downscaling (DySCo), a non-intrusive framework yielding high-resolution climate projections consistent with coarse GCM dynamics. DySCo relies on a data-driven reformulation of nudging to create dynamically paired training trajectories without intrusive GCM modifications. Using these paired trajectories, we train a dynamically and statistically consistent, two-stage operator. We evaluate the method by downscaling the Community Earth System Model v2 Large Ensemble (LENS2) in time and space towards historical reanalysis. Results show DySCo achieves superior dynamical consistency with the coarse GCM trajectories, essentially applying a minimal, causal correction to the GCM, preserving top statistical performance comparable to state-of-the-art unsupervised models.
Abstract. A key challenge in flood risk analysis is the construction of hazard events that are physically plausible yet extend beyond historical observations with appropriate frequency and spatial coherence. This is commonly addressed through large simulations of synthetic weather scenarios that sample low-likelihood, high-impact events beyond the observed record. Although popular in industrial risk-based workflows, traditional statistical approaches to synthetic weather generation can be limited in their ability to represent the full range of physically plausible variability and spatial structure. Here, we demonstrate a framework that uses an AI-based weather model as a stochastic generator of event sets suitable for flood risk assessment. We adapt the huge ensembles (HENS) approach using a Spherical Fourier Neural Operator (SFNO)-based atmospheric model combined with a diagnostic precipitation model, forming a framework termed “PrecipHENS”. This framework produces more than 1000 synthetic European winter seasons of precipitation and temperature at 0.25° resolution, with modest computational cost (using NVIDIA Earth-2 stack, 112 GPU hours on NVIDIA L40s GPUs). Using an Elbe River case study, we evaluate PrecipHENS against risk-relevant criteria, including reproduction of present-day climatology, preservation of spatial and temporal dependence, representation of extremes, and extrapolation beyond the historical record in event space. PrecipHENS reproduces key features of precipitation and temperature climatology, preserves spatial dependence, including the decay of extremal co-occurrence with distance, and generates a substantially broader diversity of extreme precipitation events than an industry-standard conditional multivariate extreme-value benchmark. Principal component analysis of extreme precipitation fields shows that PrecipHENS spans a much broader space of storm structures than the benchmark or the historical record, indicating it is able to produce previously unseen weather rather than repetition of past patterns. To assess flood risk relevance, the AI-generated weather sequences are coupled with a hydrological model. The resulting river flow simulations are consistent with observed climatology and extreme discharge behaviour, demonstrating that meteorological realism translates into physically plausible hydrological response. Together, these results show that AI-based large-ensemble weather generation can support event set construction for flood hazard and flood risk applications. More broadly, this framework provides a pathway for expanding the physically plausible sample space in applications that require robust characterisation of extremes, including risk assessment, climate-impact analysis, and storyline development.
J. Ashcroft, Alison Poulston, Marius K. Koch et al.· Natural Hazards and Earth Sy...· 0 citations
Accurately estimating rainfall distributions, from-small-to-extreme totals, is crucial for addressing various environmental challenges (e.g., flood forecasting, water resource management, disaster preparedness). Global Numerical Weather Prediction (NWP) models can provide useful rainfall estimates; yet, they often misrepresent point-scale observations from rain gauges, underestimating the frequency of small rainfall totals and extreme values. In general, finer resolutions yield more accurate representation of gauge-based climatologies. Hence, this study provides a systematic, global verification of four NWP-modelled rainfall datasets of differing resolutions (with “resolution” meaning “horizontal grid spacing”) - ERA5’s Ensemble Data Assimilation (62 km, probabilistic), ERA5’s short-range forecasts (31 km, deterministic), short-range 46r1 ECMWF reforecasts (18 km, control run), and ERA5-ecPoint (point-scale, probabilistic)—against 20 years of global rain gauge observations, assessing each dataset’s ability to represent the entire rainfall distribution. Although in very mountainous areas (e.g., the Andes) ERA5-ecPoint underestimates zero-rainfall frequency and overestimates wet tail length, it dramatically improves upon raw NWP performance in many other regions by capturing more accurately the frequency of zeros, the “growth rates” of rainfall totals, and the wet tails. Moreover, due to its probabilistic nature, ERA5-ecPoint can estimate long return periods (e.g., 1000 years) without using distribution fitting, thereby offering valuable insights into extremely rare or unprecedented events at specific locations. Such findings underscore the importance of using post-processing to enhance the local-scale validity of global NWP models. Moreover, as climate change intensifies extreme rainfall events, such post-processing becomes crucial for estimating accurate long-period rainfall climatologies, as needed for effective mitigation and resilience building, particularly in areas lacking comprehensive and reliable rain gauge records.
F. Pillosu, T. Hewson, E. Gascón et al.· Bulletin of The American Met...· 0 citations
This study investigates how uncertainties in high-resolution observation-based gridded datasets (OBGDs) influence downscaled climate projections in the Puget Sound region of the Pacific Northwest, U.S. We compare four OBGDs (gridMET, nClimGrid, Livneh, and GMFD) with station observations and identify significant disagreement in annual Frost Days. These biases influence uncertainty in three widely used bias-corrected and statistically downscaled (BSD) products (STAR-ESDM, LOCA2, NEX-GDDP-CMIP6), resulting in mid- and late-century projections that differ by up to 100% in comparisons based on the same sixteen CMIP6 models. Differences among BSD products also exceed 1°C in winter minimum temperature warming, 50 Frost Days and 30 Summer Days in areas with complex terrain. These findings emphasize that high spatial resolution does not ensure local accuracy, and reliance on a single dataset can obscure critical uncertainties. This has important implications for infrastructure and ecosystem planning, where decisions are often based on temperature thresholds. We recommend users consider multiple OBGDs and BSD products and account for known biases when using climate data for decision-making and probabilistic projections.
Graham W. Taylor, Keith W. Dixon, Liqiang Sun et al.· Journal of Applied Meteorolo...· 0 citations
Tropical Cyclones (TCs) damage assets and threaten populations globally, multiple times a year. Forecasting products from meteorological agencies across the world can help anticipate their likely path, intensity and broad regional impact. They are critical in informing safety warnings and potential evacuation measures at the county scale. These products are not designed to represent experience on the ground at a scale characteristic of individual neighborhoods (i.e. the local scale, ~10
3
m). This limits their usability for granular decision making. Local-scale simulations are achievable using full physics numerical weather prediction models, but the associated computational requirements typically allow for only a handful of deterministic simulations to be performed in real-time. This restricts their use in applications requiring probabilistic information. Recent developments from AI based weather forecasting models provide vastly more efficient TC forecasting solutions that can run simulation ensembles to provide probabilistic information in real-time. Yet these are fundamentally limited by the resolution of the data they train on, which currently fails to represent the local scale. We here introduce LiveCyc, a machine learning approach that can augment any TC track and intensity forecast with a probabilistic local-scale wind forecast. After an overview of the algorithms forming LiveCyc, we introduce an extensive dataset of historical back tests. Using this dataset, we show the value of LiveCyc in informing local-scale decision making in the days before a TC makes landfall. In particular, we demonstrate how objective cost-saving optimization can calibrate and automate the triggering of protective actions ahead of a TC event.
T. Loridan, N. Bruneau, B. Mani et al.· Bulletin of The American Met...· 0 citations
Identifying the causes of Earth's extremes is challenging because counterfactual experiments are not possible in the observed world. Data-driven causal discovery complements computationally expensive and potentially biased numerical experiments, but existing methods can struggle with undersampled, high-dimensional data and fail to recover multi-timestep, multivariate pathways leading to specific events. We introduce Tracer of Causal Evolutions in Space and Time (TraCE-ST), a probabilistic Lagrangian approach that produces event-conditioned causal trajectories in multivariate gridded data. TraCE-ST recovers known causal drivers and estimates their relative contributions in synthetic experiments and real-world extremes, including the 1991 Mount Pinatubo eruption. TraCE-ST also highlights less-studied drivers, such as orography-driven vorticity for Tropical Storm Debby (2006) and anomalous ocean-surface fluxes for the 2021 Pacific Northwest heatwave. Here, we propose causal tracking as an efficient data-driven framework for synthesizing causal evidence and generating testable hypotheses, complementing association analyses and numerical modeling while accelerating the study of high-impact events.
Jhayron S. Pérez‐Carrasquilla, J. Nichol, Vanessa Robledo et al.· 0 citations
Climate extremes in Europe are becoming increasingly frequent and severe, heightening the need for reliable regional climate information to support preparedness and adaptation. Downscaled climate model (DCM) products are widely used for this purpose. Yet, their accuracy relative to global climate models (GCMs) and observations remains uncertain, particularly for extremes. We evaluated DCM ensembles (EURO-CORDEX, NEX-GDDP-CMIP6, GDPCIR) against GCMs (CMIP5, CMIP6) and reanalysis datasets (ERA5, GMFD), using gridded observational data (E-OBS) as reference. Model accuracy was assessed for mean and extreme temperature and precipitation across Europe using complementary metrics that quantify distributional accuracy, pointwise agreement, and bias magnitude and direction. The results show that downscaling can add details to and increase the accuracy of modeled climate variables in Europe. Statistically downscaled and bias-corrected products NEX-GDDP-CMIP6 and GDPCIR generally outperform their driving GCM for mean and extreme temperature, and for mean precipitation, whereas gains for extreme precipitation are limited. Compared to E-OBS, the dynamically downscaled ensemble EURO-CORDEX exhibits some strong regional biases, notably for temperature. No single dataset performs best in all regions. Across datasets, accuracy is reduced in areas with complex terrain (mountains and coastlines). Uncertainties of the observational benchmark further complicate the evaluation. Our findings highlight both the value and limitations of climate model downscaling: while DCMs provide critical fine-scale information, their reliability is variable- and region-dependent, with extreme precipitation and conditions in complex terrain remaining particularly difficult to capture.
M. Hulkkonen, Akash Deshmukh, T. Mielonen et al.· npj Climate and Atmospheric...· 0 citations
