From climate projections to trajectory-conditioned local predictions
Under review (2026)
Abstract:
Regional climate projections are usually sharpened by constraining models’ forced responses toward observations, a step that filters out internal variability and works well mainly where the signal-to-noise ratio is high. Here we side-step this limitation with deep learning, treating internal variability not as noise to remove but as part of the future to predict and of the uncertainty to report. A U-Net trained on 290 simulations from 40 CMIP6 Earth System Models maps the 1980 to 2025 evolution of surface air temperature onto decadal-mean monthly climatologies from the 2030s to the 2090s under SSP2-4.5, keeping the forced response and internal variability together. Tested on entirely withheld model families, the forecasts are skilful and empirically calibrated out-of-sample. Within a model's own ensemble, a forecast predicts the future of the specific member it was given more accurately than the futures of that model's other members, a skill that persists to the 2090s, notably over the subpolar North Atlantic, Europe, North America and Southern Ocean. The forecasts also generalise to a next-generation eddy-rich model whose resolved physics lies outside the training distribution, supporting application to real-world data. When applied to observations, the framework reduces uncertainty in global-mean warming similarly to observational constraints, while also providing spatially coherent local predictions. This shows that our deep learning approach can estimate the distribution of future local climate, conditioned on the recent observed trajectory, encompassing uncertainty arising from both forced response and internal variability realisations. This extends observationally constrained projection toward calibrated predictions at the local scale and seasonal resolution where adaptation decisions are made.
Unravelling the role of increased model resolution in simulating surface temperature fields using explainable AI
Artificial Intelligence for the Earth Systems, In press (2026)
Abstract:
Minimizing climate model biases is essential to reduce uncertainties in future climate projections. Despite recent advances, improvements between generations of Earth System Models (ESMs) remain modest, largely due to continued reliance on subgrid-scale parametrizations. These are required because CMIP6 model resolutions are too coarse to explicitly simulate small-scale processes such as ocean mesoscale eddies and deep atmospheric convection, which strongly influence climate patterns. Recent gains in computational power have enabled higher-resolution ESMs that can resolve some of these processes, thereby reducing the need for parametrization. However, robustly detecting improved modeling with increased resolution remains challenging due to internal variability and model-to-model discrepancies. This study employs a convolutional neural network (CNN) classifier combined with explainable AI (XAI) to assess the role of resolution in simulating winter surface temperature fields in an ensemble of 17 control climate simulations with varying oceanic and atmospheric resolutions. The CNN distinguishes between ESMs of varying resolutions using temperature snapshots, while XAI identifies the regions driving these classifications, offering deeper insight into ESM behavior. The results show that ESMs with similar ocean grid resolutions are more often confused with each other by the CNN than those from different modeling centers, highlighting the central role of ocean resolution, particularly mesoscale eddies, in shaping climate simulations. Although limited to surface air temperature, the approach provides a more nuanced perspective on ESM differences and performance than traditional bias analyses. The framework can be extended to other variables and ESM features, offering a powerful tool for ESM intercomparison and evaluation.
Local temperatures rapidly depart from historical experience beyond 1.5°C
Under review (2026)
Abstract:
Local temperatures increasingly exceed values without precedent in the historical record, but when these excursions become permanent is uncertain due to the underestimated role of internal climate variability. Here we recast the problem in probabilistic terms by asking, for a given global warming level, how likely is it that annual mean local temperature will remain above its 1850-1999 maximum through 2100, given the preceding climate history. We train a neural network on 408 CMIP6 simulations under SSP2-4.5, the emissions trajectory closest to current policy commitments, and obtain calibrated probabilities skilful under out-of-sample validation. Applied to observations, regions where temperatures remain above their historical maximum are confined to just 2.2% of global land area at 1.5°C, but expand rapidly thereafter, reaching nearly half the present-day global population and 35% of aboveground biomass by 2°C, and most of humanity and nearly 80% of aboveground biomass by 3°C. Current policy trajectories will therefore commit a rapidly growing share of humanity and Earth’s ecosystems to twenty-first century local temperatures sustained above their historical maxima.
Short- to long-range climate forecasts with deep learning
Copernicus Publications (2026)
Abstract:
Uncertainty in projections of future regional climate change remains large, driven by structural differences among Earth System Models and the influence of internal climate variability. Existing uncertainty-reduction approaches, including emergent constraints and Bayesian variants, primarily focus on forced climate responses derived from simple aggregate metrics, thereby requiring strong assumptions and exploiting only low-dimensional climate information. Here we propose a data-driven deep-learning framework that directly forecasts spatially and monthly resolved decadal mean climatologies of surface temperature anomalies from the 2030s to the 2090s, using only recent monthly trajectories spanning 1980-2025. The training ensemble contains 265 historical+SSP2-4.5 simulations, distributed across 40 ESMs from 25 different families (i.e., modelling centers) over which the cross validation is performed. The architecture couples pluri-annual to multi-decadal temporal convolutions with a spatial U-Net encoder-decoder and is evaluated on CMIP6 simulations using a leave-one-model-family-out cross-validation (LOMFO-CV) design to ensure generalisation across separately developed ESMs. Predictive uncertainty is quantified via LOMFO-CV errors, yielding conservative and reliable ranges that incorporate irreducible internal variability and systematic model shifts.To further evaluate the predictive capacity beyond the CMIP6 distribution, we evaluated the network on historical+SSP2-4.5 simulations from a recent HadGEM3-GC5 model hierarchy developed within the European Eddy-Rich ESMs (EERIE) project, the European contribution to HighResMIP2 for CMIP7. In particular, the eddy-rich GC5-HH configuration explicitly simulates mesoscale ocean dynamics that are absent in CMIP6-type models, providing a rigorous test of generalisation to richer and more realistic physical representations. Despite these substantial differences, the network successfully reproduces warming trajectories and future climate patterns for all three model configurations (GC5-LL, GC5-MM, GC5-HH), with forecast errors largely contained within empirically calibrated uncertainty bounds from the LOMFO-CV, both globally and locally. These results, notably for GC5-HH and its more realistic physics, strengthens confidence in the applicability of the framework to real-world data.When applied to observations, the extracted end-of-century global-mean surface temperature and its uncertainty range are consistent with prior estimates from Bayesian frameworks. At local scales, the network reduces uncertainty by 40% (2030s) to 30% (2090s) on average, and by up to 75% in some regions for all future decades. Importantly, these uncertainty estimates account not only for uncertainty in the forced response (as emergent constraint methods do), but also for errors associated with predicting different realisations of internal variability, providing a physically meaningful reduction of local and global climate uncertainty.
Reconstruction of last millennium sea surface temperature on 1° grid using a random forest algorithm
Global and Planetary Change 258 (2026) 105279
Abstract:
Climate models and theoretical evidence show that the ocean drives climate from sub-decadal to centennial timescales through a variety of processes and their interactions. The range of direct climate observations, however, is too short to understand the exact role of the ocean in shaping observed and future climate variability on top of anthropogenic climate change. In the present study, we use a large set of paleoclimate records combined with a random forest algorithm to reconstruct a gridded dataset of sea surface temperatures since 850 C.E. to provide a better framework for the study of ocean surface variability. In line with modeling and paleodata studies, our reconstruction suggests that natural climate forcings have importantly influenced the last millennium climate variability. Our reconstruction also suggests that North Atlantic SST multidecadal variability influences Pacific SST on decadal timescales. However, the latter result is shown to be strongly dependent on background climate conditions. This new reconstruction offers a useful resource for testing the capabilities of climate models to reproduce the linkages between Atlantic and Pacific as well as the response to external forcings.