Abstract Adapting to the changing climate requires downscaling of low-resolution climate information into products that are local-scale, accurate, and capable of capturing extremes. Most current forms of downscaling fall into two categories: statistical downscaling, which is computationally efficient but poor at downscaling extremes, and dynamical downscaling, which can more effectively capture extremes and complex patterns without assumptions of stationarity but is computationally demanding. Deep learning (DL) models offer a potential bridge between these two categories; by training a DL model on high-resolution output of dynamical downscaling, they can learn to efficiently emulate the regional climate model. This study investigates the strengths and weaknesses of using a stochastic generative adversarial network (GAN) to downscale climate variables. First, we show that the stochastic GAN framework successfully downscales temperature, specific humidity, and precipitation, although misses some fine-scale details for humidity. We also find that the stochastic framework substantially improves the downscaling of extreme precipitation. Second, we develop a multihead GAN and show that multivariate downscaling can improve dependence structures between downscaled variables while maintaining high accuracy output. Finally, we test the portability of the GAN framework to a new location with a different climate and show that while the GAN performs well for wind, temperature, and humidity, it is less accurate for precipitation due to mismatches between the low- and high-resolution data. These results provide important progress toward operational GAN-based downscaling.

Abstract To enable the study of the climatology of convective and stratiform precipitation in tropical cyclones, this work develops a convective–stratiform precipitation-type classification model for the Global Precipitation Measurement satellite Microwave Imager (GMI) from the Tropical Cyclone Precipitation, Infrared, Microwave, and Environmental Dataset. The model uses the random forest classifier and takes as input the various brightness temperature observations from the GMI and calculated texture information. The optimal model setup is found through experimenting with the minimum number of samples in a leaf node and model calibration to prioritize classification accuracy. The final model is the sigmoid-calibrated model with a minimum number of samples in a leaf node of four. Overall, the model does well at capturing the general precipitation structure of tropical cyclones. Model deficiencies are likely due to issues with nonuniform beam filling. Finally, a dropout experiment is performed to investigate the most important predictors. The most important predictors are the 36.64- and 89.0-GHz brightness temperatures and texture information. While texture information are important, brightness temperature observations at observing frequencies of less than 89.0 GHz have a more significant impact toward model skill in classifying precipitation type from passive microwave observations.

Abstract We discuss the utility of applying clustering as a preprocessing step for identifying subseasonal to seasonal forecasts of opportunity of coastal sea level using convolutional neural networks (CNNs). Clustering leverages potential covariance due to modes of climate variability among points along the same coastline or in the same ocean basin. To evaluate the utility of clustering for reliably identifying forecasts of opportunity, we compare CNNs trained to predict sea level probability distributions in three ways: over the whole northeast Pacific Coast simultaneously, over predetermined clusters within this coastline, and at individual grid points near tide gauges. All CNN prediction tasks (Whole Coast, Cluster, Point) outperform climatology by a similar margin at week 3, when the entire test set is used to evaluate CNN skill. However, when comparing the skill of each task’s 20% most confident predictions, we find the skill of the Cluster and Point tasks to be on par with each other and substantially more skillful than the Whole Coast task. Of the Cluster and Point tasks, the Cluster task represents all grid points in the northeast Pacific Coast with minimal tunable parameters. Throughout this exercise, we learned that clustering grid points as a preprocessing step is the preferred approach between the three for making S2S predictions of coastal sea level. Significance Statement When building CNNs to make predictions for geophysical variables, researchers need to decide how to partition the domain of interest. This paper explores three ways to partition the northeast Pacific Coast to make subseasonal to seasonal predictions of sea level. We find that preclustering the points and training CNNs to predict those clusters yield skillful predictions which can cover the entire domain of interest with a minimal number of CNNs. This work underscores the importance of clustering as a method for dimensionality reduction in geophysical systems and implies that clustering based on spatial or temporal covariability can lead to increased predictive skill.

Abstract Accurate and computationally efficient modeling of ocean currents is important for understanding and monitoring the dispersion of marine pollution. Passive floating Lagrangian drifters, equipped with global positioning systems (GPSs), are a low-cost approach to tracking surface flow. However, assimilating drifter observations into comprehensive ocean models remains challenging, and these models require calibration of parameters and are often too computationally intensive for real-time or operational use. We introduce a Lagrangian data assimilation method that leverages machine learning techniques to enable computationally efficient ocean flow modeling. We first use reanalysis data to train an autoencoder to build a reduced-order model that extracts low-dimensional representations of the high-dimensional flow field. The decoder serves as a generative model, producing plausible flow fields from samples drawn from the autoencoder latent space. We then apply Bayesian optimization to minimize a cost function, defined in the low-dimensional latent space, that measures the discrepancy between observed trajectories and model-generated trajectories. Our approach is demonstrated to efficiently reconstruct regional ocean flow from sparse drifter data, with results obtained within a few minutes using a standard graphics processing unit (GPU). The method can be used to identify the minimum number of drifters needed for accurate flow modeling and to determine optimal drifter launch locations. We evaluate the method on a synthetic Kolmogorov flow, as well as a realistic flow field in the Massachusetts and Cape Cod Bays. Significance Statement Tracking and predicting ocean currents are important for monitoring how pollutants, such as wastewater or runoff, spread through the ocean. Lagrangian drifters provide valuable, low-cost observations of surface flow, but current methods for using these observations in models are often too computationally expensive for real-time applications. We develop a new data assimilation approach that uses machine learning to reconstruct ocean flow efficiently from drifter observations. By combining an autoencoder-based reduced-order model with Bayesian optimization, the method can estimate regional flow patterns in just minutes. Our results show that this method can provide rapid and accurate reconstructions of ocean current patterns in a region of interest, demonstrated in both an idealized simulation and in the Massachusetts and Cape Cod Bays.

Abstract Machine learning weather prediction (MLWP) models have demonstrated remarkable potential in delivering accurate forecasts at substantially reduced computational cost compared to traditional numerical weather prediction (NWP) systems. However, challenges remain in ensuring the physical consistency of MLWP outputs, particularly in deterministic settings. This study presents FastNet, a graph neural network (GNN)-based global prediction model, and investigates the impact of alternative loss function designs on improving the physical realism of its forecasts. We explore three key modifications to the standard mean-square error (MSE) loss: 1) a modified spherical harmonic (MSH) loss that penalizes spectral amplitude errors to reduce blurring and enhance small-scale structure retention, 2) inclusion of horizontal gradient terms in the loss to suppress nonphysical artifacts, and 3) an alternative wind representation that decouples speed and direction to better capture extreme wind events. Results show that while the MSH and gradient-based losses alone may slightly degrade root MSE (RMSE) scores, when trained in combination, the model exhibits very similar MSE performance to an MSE-trained model while at the same time much improved spectral fidelity and physical consistency. The alternative wind representation further improves wind speed accuracy and reduces directional bias. Collectively, these findings highlight the importance of loss function design as a mechanism for embedding domain knowledge into MLWP models and advancing their operational readiness.

Abstract Hail poses a substantial economic threat in Switzerland, where both property and agricultural insurers report it as the leading source of storm-related losses in recent years. Accurate detection and sizing are critical for effective risk assessment and warning systems. This study presents a machine learning approach to predict hail presence and size at the ground for storm objects. We used Swiss C-band polarimetric radar data and a large, crowdsourced hail report database to train two random forest models on an extensive set of radar-derived predictors: a binary model for hail detection and a regression model for size estimation. Model development followed a rigorous pipeline based on repeated recursive predictor elimination to identify final models that balance performance, robustness, stability, and meteorological interpretability. The final hail detection model outperforms the operational probability of hail (POH) baseline (Peirce skill score 0.91 vs 0.81) and relies on a stable, physically coherent set of nine predictors. The final hail size regression model achieves better error statistics than the operational maximum expected severe hail size (MESHS) algorithm [root-mean-square error (RMSE) 7.5 vs 13.9 mm]; however, it substantially underpredicts large hail (≥32 mm), a bias likely exacerbated by the dataset’s strong imbalance toward smaller hail sizes. Overall, this study demonstrates that a data-driven approach with a focus on model stability can improve hail detection and size estimation for storm objects compared to currently operationally used radar-based hail detection algorithms. Importantly, it delivers robust and physically interpretable models rather than opaque “black boxes,” making them promising candidates for further validation and application. Significance Statement Hail causes more insurance losses in Switzerland than any other storm-related hazard. To improve warnings and risk assessment, we need to know exactly how big the hail is, but current radar techniques struggle with this precision. We used machine learning to combine radar data with over 270 000 hail reports from the public. Our method detects hail more reliably than current operational systems and reduces average size estimation errors by nearly half. While predicting rare, massive hailstones remains difficult, this study shows that a small set of distinct radar patterns can effectively link storm structure to hail size, providing a robust foundation to inform future operational hail detection systems.

Abstract Atmospheric predictability research has long held that rapid error growth at small spatial scales imposes an intrinsic limit of roughly 2 weeks on deterministic weather forecast skill. We challenge this limit using GraphCast, a machine learning weather model, by optimizing initial conditions for twice-daily forecasts spanning 2020. This approach yields an average error reduction of 86% at 10 days relative to control forecasts from reanalysis initial conditions, with skill lasting beyond 30 days. Mean optimal initial-condition perturbations reveal large-scale, spatially coherent corrections primarily reflecting an intensification of the Hadley circulation. Forecasts using GraphCast-optimal initial conditions in the Pangu-Weather model achieve a 21% error reduction, peaking at 4 days, indicating that analysis corrections reflect adjustments that target both model and analysis error. These results demonstrate the existence of initial conditions producing skillful deterministic forecasts far beyond 2 weeks. Whether such initial conditions can be identified in real time for improving operational weather forecasts remains a topic of future research. Significance Statement Our results contest the long-standing view that rapid error growth at small spatial scales fundamentally limits reliable weather forecasts to about 2 weeks. By applying machine learning–based optimization to atmospheric initial conditions, we achieve skillful deterministic hindcasts to at least 33 days—more than twice the conventional limit. Validating these optimized initial conditions in an independent weather model leads to significant but smaller improvements, highlighting the importance of model error in the procedure. On average, the optimal corrections reveal physically consistent large-scale patterns that align with an intensified tropical overturning circulation. These results demonstrate the existence of initial conditions that considerably extend the current established limit of predictability. Real-time determination of such initial conditions for operational forecasting remains an open question.

Abstract Uncertainty in reference to machine learning (ML) systems is notable across many scientific domains due to its influence on critical decision-making processes. But further attention is needed when using ML-uncertainty quantification (UQ) approaches for rare and impactful environmental events. Three ML-UQ multimodel ensemble (MME) methods using three different loss functions—mean-square error (MSE), negative log-likelihood (NLL), and continuous ranked probability score (CRPS)—were compared to understand which methods best estimate uncertainty of water temperature predictions at different lead times (12, 48, 96 h). The models were designed and implemented to improve cold-stunning event advisories and decision-making in South Texas (periods when water temperatures reach threatening levels for wildlife). A 10-fold cross-validation framework was used to compare deterministic and probabilistic skill among the UQ methods during cold seasons and sub-12°C cases. UQ methods were further compared on unseen data during rare, high-impact cold water conditions, including the most recent and the most impactful event on record. Statistical analyses were performed to identify any significant differences among UQ methods. Although MSE-based ML predictions have historically supported cold-stunning advisories, they offer limited reliability in uncertainty estimates during critical conditions. In contrast, both CRPS-MME and NLL-MME provide significantly better deterministic and probabilistic performance, with CRPS-MME exhibiting superior performance at shorter lead times and NLL-MME performing best at longer lead times under more typical cold season conditions. However, CRPS-MME outperformed all UQ methods for the most extreme cold-stunning event since 1989, highlighting its potential for high operational utility for impactful cold-stunning events along the Texas coast. Significance Statement This study compares and evaluates how effectively different machine learning uncertainty quantification (ML-UQ) ensemble approaches estimate uncertainty of water temperature predictions for impactful cold-stunning events. Methods explicitly trained to characterize predictive uncertainty produced more reliable uncertainty estimates than ensembles composed of deterministic models, although predictive performance varied across forecast lead times and temperature conditions. These results highlight both the benefits and limitations across ML-UQ methods for cold-stunning events and provide guidance for selecting ML-UQ methods for operational forecasting systems. This probabilistic insight can support more informed and timely decision-making for cold-stunning event management efforts. This work also established a foundation for future user-centric research examining how probabilistic information can be communicated and applied effectively within real-world, time-sensitive environmental decision-making processes.

Abstract Ensemble weather predictions can be critical toward making informed decisions for operations that require risk assessment. The traditional approach toward creating an ensemble often involves making slight perturbations to initial conditions and model numerical solvers and physics to account for forecast uncertainty. This study instead uses diffusion, a form of generative artificial intelligence (AI), to make ensemble forecasts consisting of 1000 members from single-member deterministic dynamical Western Weather Research and Forecasting (West-WRF) Model forecasts over the North Pacific. After it has been trained, the diffusion ensemble can make predictions at high speeds and at a low computational expense while improving the skill of medium-range integrated water vapor transport (IVT). It has high performance when evaluated by binned ranked histograms and outperforms operational ensemble forecasts [Global Ensemble Forecast System (GEFS) and European Centre for Medium-Range Weather Forecasts (ECMWF)] in continuous ranked probability score, binned spread to root-mean-square error ratio, reliability diagrams, and Brier skill scores. The diffusion ensemble further demonstrated utility in representing real outcomes of extreme destructive events with realistic high-quality images. Significance Statement This work explores a novel method of creating probabilistic weather forecasts. By harnessing diffusion, a form of generative artificial intelligence, a large ensemble of medium-range forecasts of integrated water vapor transport was created using a deterministic forecast as input. The diffusion ensemble had better performance than several operational weather models based on a range of probabilistic skill metrics.

Abstract Fire radiative power (FRP) detection of wildland fires is important in calculating a variety of variables, such as smoke emissions, which are necessary for fire weather and air quality simulations. FRP is derived from satellite observations of radiative energy emitted by fires. In this work, FRP values collected by geostationary and polar-orbiting satellites in 2018 are used in combination with weather forecast model outputs to derive a next-hour forecasted hourly FRP value unique to that model grid point. Random forest (RF) models were trained and applied to multiday burning fires with inputs including day-before average FRP satellite values, hour of day, location, temperature, wind speed, and relative humidity to produce a next-hour FRP value on the Rapid Refresh (RAP) model resolution. Models were trained on separate satellite FRP inputs given the difference in satellite sensor resolutions: one used Geostationary Operational Environmental Satellite (GOES) FRP inputs and the other used a collection of polar-orbiting satellite FRP values. Overall, mean absolute errors (MAEs) for all RF models were lower than those for the current persistence approach method used by the High-Resolution Rapid Refresh (HRRR). The main lesson learned from this paper is that a simple RF model is a good alternative to the persistence method and provides a uniquely forecasted hourly FRP value based on weather and time/location. Significance Statement This work introduces a type of machine learning model, random forest (RF), to the current method in which hourly fire radiative power (FRP) is modeled using a day-before average FRP value and a hour of day value. By using an RF, numerical weather model information is used as an input with both the day-before average FRP value and the time of day value to create a grid point and weather-specific hourly FRP forecast.