Conformal prediction (CP) can yield statistically valid prediction intervals for any regression model, with no model modifications and small computational costs. To assess its practical value, we apply conformal methods to quantify uncertainty in machine learning emulators of six microphysical process rates (MPRs). MPRs describe small-scale processes in atmospheric clouds such as precipitation formation and aerosol–cloud interactions and help understand weather and climate. The emulators are trained on simulation output from the ICOsahedral Nonhydrostatic (ICON) model in a limited-area numerical weather prediction configuration. We compare split CP for deterministic emulators with conformalized quantile regression (CQR) for quantile regression (QR) emulators. Both CP methods yield well-calibrated and sharp prediction intervals on average, but CQR provides more consistent intervals across several orders of magnitude, making it preferable for the uncertainty quantification of climate variables.

The “differentiability gap” presents a primary bottleneck in Earth system deep learning: Since models cannot be trained directly on non-differentiable scientific metrics and must rely on smooth proxies (e.g., MSE), they often fail to capture high-frequency details, yielding “blurry” outputs. We develop a framework that bridges this gap using two different methods to deal with non-differentiable functions: The first is to analytically approximate the original non-differentiable function into a differentiable equivalent one; the second is to learn differentiable surrogates for scientific functionals. We formulate the analytical approximation by relaxing discrete topological operations using temperature-controlled sigmoids and continuous logical operators. Conversely, our neural emulator uses Lipschitz-convolutional neural networks to stabilize gradient learning via: (1) spectral normalization to bound the Lipschitz constant and (2) hard architectural constraints enforcing geometric principles. We demonstrate this framework’s utility by developing the Minkowski image loss, a differentiable equivalent for the integral-geometric measures of surface precipitation fields (area, perimeter, and connected components). Validated on the EUMETNET OPERA dataset, our constrained neural surrogate achieves high emulation accuracy, completely eliminating the geometric violations observed in unconstrained baselines, which generate physically impossible precipitation fields in up to 7.8% of cases. However, applying these differentiable surrogates to a deterministic super-resolution task reveals a fundamental trade-off: While strict Lipschitz regularization ensures optimization stability, it inherently over-smooths gradient signals, restricting the recovery of highly localized convective textures. This work highlights the necessity of coupling such topological constraints with stochastic generative architectures to achieve full morphological realism.

In recent years, deep learning has transformed data processing, justifying its use in the current study. These models consist of multiple layers that calibrate weights to optimize accuracy. By mimicking neural processes, these networks facilitate high-performance predictive modeling. This study focused on the application of convolutional neural networks for forecasting Sargassum seaweed arrivals in Guadeloupe, using varied climatic data such as winds, surface currents, satellite images, and observations of seaweed arrivals on the Guadeloupe coast. We tested two different approaches: classification models for forecasting up to 14 days and zone regression models for forecasting up to 14 days, also indicating a probability of algae arrival. The results showed that the classification models achieved a mean accuracy of 94.75%, a mean specificity of 97.42%, and a mean sensitivity of 98.78%. Zone regression models achieved a mean accuracy of around 90%, a mean specificity of around 97%, and a mean sensitivity of around 85%, with a mean MAE of 0.13 and a mean RMSE of 0.18. These models perform better than the decision tree proposed in the state of the art for the same problem.

Recent advances in visual-language models (VLMs) have led to significant improvements in a plethora of complex multimodal tasks like image captioning, report generation, and visual perception. However, generating text from meteorological data is highly challenging because the atmosphere is a chaotic system that is rapidly changing at various spatial and temporal scales. Given the complexity of atmospheric phenomena, it is critical to verifiably quantify the effectiveness of existing VLMs on weather forecasting data. In this work, we present SynopticBench, a high-quality dataset consisting of 1,367,041 text samples of Area Forecast Discussions created by the National Weather Service over the continental United States paired to images of 500 mb geopotential height, 2 m temperature, and 850 mb wind velocity in weather forecasts. We also present Synoptic Phenomena Alignment and Coverage Evaluation (Space), a novel evaluation framework that can be used to effectively estimate the quality of text descriptions of synoptic weather phenomena. Extensive experiments on generating forecast discussions using state-of-the-art VLMs show the sensitivity of existing evaluation metrics in this domain and enable further exploration into synoptic weather and climate text generation.

Next-generation high-resolution (km-scale) climate models promise unprecedented accuracy in climate projections, but realizing their potential requires robust methods to quantify how well simulations align with real-world observations. Average-based metrics conventionally used for climate model evaluation ignore the physics encoded in the fine-scale structures of km-scale simulations. To overcome this limitation, we propose a novel, statistically principled evaluation methodology based on the likelihood function of a generative image model. Our method provides a continuous similarity metric derived from the likelihood distribution of observation and simulation snapshots, which can redefine the evaluation, intercomparison, and parameter tuning of high-resolution climate models. We demonstrate the applicability and interpretability of this method by evaluating convective clouds simulated by two state-of-the-art global km-scale models, using their outgoing infrared radiation fields. This work establishes a scalable pathway toward observation-based evaluation of next-generation climate simulations.

Multi-model ensembles are widely used in climate science, yet Coupled Model Intercomparison Project Phase 6 (CMIP6) models are neither independent nor equally skillful. We present a framework for interpretable ensemble learning that improves prediction while making model contributions explicit. Using monthly CMIP6 near-surface temperature and precipitation (1948–2014) against ERA5, we compare arithmetic multi-model ensemble (AMME) averaging with Ridge regression, random forest, and a continuous gating mixture that adaptively interpolates between them. On held-out test years (2006–2014), gating achieves the best temperature performance and improves precipitation distributional behavior relative to both AMME and individual learners. We trace these gains back to individual models and space: dominant-cluster and dominant-model maps show that gating reallocates trust in regionally coherent patterns rather than collapsing to a single best model, while entropy-based diagnostics quantify where trust is concentrated versus distributed. These analyses provide a practical pathway for transparent ensemble learning that supports both predictive skill and scientific accountability.

Subseasonal forecasts provide valuable guidance for decision-making but remain limited by systematic biases and coarse spatial resolution, particularly over regions with complex orography. Quantile mapping (QM) is widely used for post-processing because it enables effective bias correction and downscaling, yet it applies a distributional correction and therefore neglects the case-specific dependence between forecasts and observations. To move toward a more integrated approach, we explore a diffusion-based generative framework for simultaneous probabilistic bias correction and high-resolution downscaling, with lead time explicitly included as a conditioning variable. We employ a conditional denoising diffusion model (DM) trained on high-resolution gridded observations (2 km) and conditioned on coarse-resolution ensemble-mean subseasonal forecasts (36 km) from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS) subseasonal system. We focus here on a univariate temperature model and on training strategies. By learning conditional distributions of observations given forecasts and lead time, diffusion-based generative models enable case-dependent corrections that exploit forecast–observation relationships. Two training configurations are examined to assess how lead-time conditioning and data availability influence the learning of biases. Results show that the DM recovers realistic fine-scale temperature patterns and achieves meaningful lead-time-dependent bias correction across subseasonal lead times, even when trained on relatively limited datasets. However, compared to QM, the diffusion-based forecasts exhibit lower skill in probabilistic metrics such as the continuous ranked probability score (CRPS) and discrimination for extremes. Overall, this work represents a first step toward lead-time-aware subseasonal post-processing using DMs and highlights directions for future methodological development.

General circulation models (GCMs) are widely used for future climate projections, but their coarse spatial resolution and systematic biases limit their direct use for impact studies. This limitation is particularly critical for wind-related applications, such as wind energy, which require spatially coherent, multivariate, and physically plausible near-surface wind fields. Classical statistical downscaling and bias correction methods partly address this issue. Still, they struggle to preserve spatial structure, intervariable consistency, and robustness under climate change, especially in high-dimensional settings. Recent advances in generative machine learning offer new opportunities for downscaling and bias correction, eliminating the need for explicitly paired low- and high-resolution datasets. However, many existing approaches remain difficult to interpret and challenging to deploy in operational climate impact studies. In this work, we apply SerpentFlow, an interpretable, generative, domain alignment framework, to the multivariate downscaling and bias correction of wind variables from GCM outputs. This is a method that generates low-resolution/high-resolution training data pairs by separating large-scale spatial patterns from small-scale variability. Large-scale components are aligned across climate model and observational domains. Conditional fine-scale variability is then learned using a flow-matching generative model. We apply the approach to multiple wind variables downscaling, including average and maximal wind speed, zonal and meridional components, and compare it with widely used multivariate bias correction methods. Results show improved spatial coherence, intervariable consistency, and robustness under future climate conditions, highlighting the potential of interpretable generative models for wind and energy applications.

Wildfires are one of the most devastating natural disasters with wide impacts across economic sectors and society in Europe. It is, however, unclear how wildfire risk will evolve in the future under different climate scenarios, as it depends on future climate change scenarios as well as on land use and population changes. The aim of the study is to study end-of-century wildfire risk in Southern Europe, under different Shared Socioeconomic Pathways. A data-driven model is trained using historical data and a combination of other data sources representing hydro-meteorological variables, land cover, human activities, and topography. The best-performing model is the XGBoost model with an AUC of 0.93 and F1 score of 0.85. Using SHapley Additive exPlanations values, we identify the most influential predictors. The trained model is run daily using future climate projections from the Canadian Regional Climate Model version 5 under the SSP1–2.6 and SSP3–7.0 scenarios, including explicit changes in population and land cover. Future projections confirm that the wildfire risk will increase by the end of the century (2081–2100) in the SSP3–7.0, with limited changes in the SSP1–2.6 scenario. The results further show that including future land cover and population changes modulates the projected wildfire risk in scenario-dependent ways, highlighting the need to jointly account for climatic and socio-environmental drivers when assessing future wildfire risk.

Machine learning-based weather forecasting models now surpass state-of-the-art numerical weather prediction systems, but training and operating these models at high spatial resolution remains computationally expensive. We present a modular framework that decouples forecasting from spatial resolution by applying learned generative super-resolution as a post-processing step to coarse-resolution forecast trajectories. We formulate super-resolution as a stochastic inverse problem, using a residual formulation to preserve large-scale structure while reconstructing unresolved variability. The model is trained with flow matching exclusively on reanalysis data and is applied to global medium-range forecasts. We evaluate (i) design consistency by re-coarsening super-resolved forecasts and comparing them to the original coarse trajectories and (ii) high-resolution forecast quality using standard ensemble verification metrics and spectral diagnostics. Results show that super-resolution preserves large-scale structure and variance after re-coarsening, introduces physically consistent small-scale variability, and achieves competitive probabilistic forecast skill at 0.25° resolution relative to an operational ensemble baseline while requiring only a modest additional training cost compared with end-to-end high-resolution forecasting.

Biomass burning aerosols (BBAs) from Southern Africa seasonally overlie the semi-permanent South-East Atlantic (SEA) stratocumulus deck, impacting the region’s energy budget through complex aerosol-cloud-radiation-meteorology interactions. Climate model intercomparison initiatives, like the Aerosol Comparisons between Observations and Models (AeroCom), have highlighted the large inter-model variability for BBA radiative effects, especially over the SEA, due to parameterization of emission modeling and smoke properties. Observational constraints are needed to reduce these uncertainties, but correlative observational studies are typically affected by confounding meteorological influences. We propose a physically informed statistical approach, based on causal graphs applied to satellite observations, to disentangle BBA influences on shortwave radiation over the SEA and identify the main sources of statistical biases plaguing observational studies. We find that, during the fire season, BBAs cause a regional shortwave cooling of −2.5 W m−2, which can be decomposed into equal contributions from three physical pathways: aerosol-radiation interactions (ARI), adjustments to ARI, and aerosol-cloud interactions (ACI). We also perform ablation experiments with graph variants to investigate the main sources of confounding—like large-scale winds, humidity-biased retrievals, or spatial aggregation of data—and show that they result in biased radiative effect estimates (between −50% and +15%). Once free of such biases, our derived causal estimates of smoke radiative effects can be used as observational constraints to improve climate models.

Machine learning models can represent climate processes that are nonlocal in horizontal space, height, and time, often by combining information across these dimensions in highly nonlinear ways. While this can improve predictive skill, it makes learned relationships difficult to interpret and prone to overfitting as the extent of nonlocal information grows. We address this challenge by introducing data-driven integration kernels, a framework that adds structure to nonlocal operator learning by explicitly separating nonlocal information aggregation from local nonlinear prediction. Each spatiotemporal predictor field is first integrated using learnable kernels (defined as continuous weighting functions over horizontal space, height, and/or time), after which a local nonlinear mapping is applied only to the resulting kernel-integrated features and optional local inputs. This design confines nonlinear interactions to a small set of integrated features and makes each kernel directly interpretable as a weighting pattern that reveals which horizontal locations, vertical levels, and/or past timesteps contribute most to the prediction. We demonstrate the framework for South Asian monsoon precipitation using a hierarchy of neural network models with increasing structure, including baseline, nonparametric kernel, and parametric kernel models. Across this hierarchy, kernel models achieve near-baseline performance with far fewer trainable parameters, indicating that much of the relevant nonlocal information can be captured through a small set of interpretable integrations when appropriate structural constraints are imposed.

Accurate and high-resolution estimation of land surface temperature (LST) is crucial in estimating evapotranspiration, a measure of plant water use and a central quantity in agricultural applications. In this work, we develop a novel statistical method for downscaling LST data obtained from NASA’s ECOSTRESS mission, using high-resolution data from the Landsat 8 mission as a proxy for modeling agricultural field structure. Using the Landsat data, we identify the boundaries of agricultural fields through edge detection techniques, allowing us to capture the inherent block structure present in the spatial domain. We propose a block-diagonal Gaussian process (BDGP) model that captures the spatial structure of the agricultural fields, leverages independence of LST across fields for computational tractability, and accounts for the change of support present in ECOSTRESS observations. We use the resulting BDGP model to perform Gaussian process regression and obtain high-resolution estimates of LST from ECOSTRESS data, along with uncertainty quantification. Our results demonstrate the practicality of the proposed method in producing reliable high-resolution LST estimates, with potential applications in agriculture, urban planning, and climate studies.

While deep learning (DL) methods have expanded the approach to solving state estimation problems in data assimilation (DA), a robust approach requires extending them for uncertainty quantification. In this work, we propose a method to represent location-based uncertainty through random, spatiotemporally coherent displacements in physical space, which are used to sample states in the vicinity of a given reference state. We integrate this strategy into an end-to-end variational DA framework, termed 4DVarNet-LU, and compare its performance with a UNet baseline, demonstrating improved reconstruction skill through the explicit use of prior state information. Trained deterministically in output and stochastically in input, we demonstrate that at inference, we see improvement in skill when using information from initial conditions. Using observation system simulation experiments based on a single-layer quasi-geostrophic model with sparse pseudo-nadir altimetry observations, we evaluate reconstruction mean and ensemble skill. Results show that 4DVarNet-LU consistently outperforms the UNet baseline in terms of root mean square error (RMSE) and continuous ranked probability score (CRPS), with increased robustness to initial condition bias and improved temporal stability. While the method does not propose to account for all different uncertainty sources in DA, it provides a structured way for modeling and training DL models for dealing with location-based uncertainty. Future work may extend this approach to account for additional sources of uncertainty and end-to-end ensemble generation.

Effective adaptation and mitigation strategies for climate change require high-resolution projections to inform strategic decision-making. Conventional global climate models, which typically operate at resolutions of 150 to 200 km, lack the capacity to represent essential regional processes. IPSL-AID is a global to regional downscaling tool based on a denoising diffusion probabilistic model designed to address this limitation. Trained on ERA5 reanalysis data, it generates 0.25° resolution fields for temperature, wind, and precipitation using coarse inputs and their spatiotemporal context. It also models probability distributions of fine-scale features to produce plausible scenarios for uncertainty quantification. The model accurately reconstructs statistical distributions, including extreme events, power spectra, and spatial structures. This work highlights the potential of generative diffusion models for efficient climate downscaling with uncertainty estimation. Climate downscaling generally comprises two components—super-resolution and bias correction. This work focuses on the former.

Bias adjustment is standard preprocessing for climate model output used in crop-impact and damage-assessment work. Most validation studies still emphasize mean temperature, even though agricultural applications often depend on nonlinear heat metrics built from daily extremes. We evaluate how bias adjustment propagates into Extreme Degree Day (EDD) metrics, which crop-yield and damage functions use to summarize heat stress above specified temperature cutoffs. Using ERA5 reanalysis as reference, we compare EDD from raw and bias-adjusted CMIP6 models (GFDL-ESM4, MPI-ESM1-2-HR, and EC-Earth3) across European ADM1 regions for 1994–2014. We compute EDD at seven thresholds (0, 4, 8, 12, 28, 30, and 32 °C), aggregate it with harvested-area weights for seven crops under rainfed and irrigated management, and sum it over crop-specific phenological windows. Bias adjustment reduces mean absolute error at low and moderate thresholds (EDD $ \le $ 12 °C) by 37–86 °C$ \cdot $days at EDD-0 and 29–64 °C$ \cdot $days at EDD-8. At extreme thresholds (EDD $ \ge $ 28 °C), $ \Delta $Skill is small on average (1–4 °C$ \cdot $days at EDD-28), and bias adjustment increases error in about 40% of evaluated combinations. The split between helpful and unhelpful bias adjustment follows the EDD threshold more closely than model identity, which has implications for impact studies that treat bias-adjusted data as uniformly preferable.