Tapio Schneider

The latest Earth System Models (ESMs) are much more complex than previous generations, and are better at reproducing aspects of the current observed climate. However, in general, the latest ESMs do not reproduce the historical record of global warming with as much fidelity as the previous generation of models. We hypothesise that this is because the techniques used to calibrate ESMs have not kept track with the sophistication of the models, a phenomenon that we term the calibration deficit. Here we summarize the current state of the art in ESM calibration and discuss key challenges and opportunities introduced by Machine Learning and Artificial Intelligence. We outline how hybrid approaches that make use of process-based models for out-of-sample climate projections, but use ML/AI approaches for model calibration, have huge potential to revolutionise Earth System Modelling.

Effective climate risk assessment is hindered by the resolution gap between coarse global climate models and the fine-scale information needed for regional decision making. Existing downscaling methods struggle to bridge this gap: physics-based methods are too computationally expensive for large ensemble sampling of extreme events, while traditional statistical methods fail to capture multivariate spatiotemporal dependencies crucial to compound risk estimation. Furthermore, current machine learning approaches rely on temporally aligned training pairs, which are unavailable for free-running climate projections. Here we introduce GenFocal, an AI framework that generates statistically accurate, fine-scale weather from coarse climate projections without requiring paired simulated and observed events during training. GenFocal synthesizes complex and long-lived hazards, such as heatwaves and tropical cyclones, even when they are not well represented in the coarse climate projections. It also samples high-impact, rare events more accurately than leading methods. By translating large-scale climate projections into actionable localized information, GenFocal provides a powerful paradigm to improve climate adaptation and resilience strategies.

Ensemble Kalman methods were initially developed to solve nonlinear data assimilation problems in oceanography but are now popular in applications far beyond their original use cases. Of particular interest is climate model calibration. As hybrid physics and machine‐learning models advance, the number of parameters and complexity of parameterizations in climate models will continue to grow. To fully realize these advances, we must move from laborious hand‐tuning to calibration‐driven model development in rapid iteration cycles. Thus, robust calibration of these parameters plays an increasingly important role. We focus on learning climate model parameters by minimizing the misfit between modeled and observed climate statistics in an idealized setting. Ensemble Kalman methods are a natural choice for this problem because they are derivative‐free, scalable to high dimensions, and robust to noise caused by statistical observations. Given the many variants of ensemble methods proposed, an important question is: Which ensemble Kalman method should be used for climate model calibration ? To answer this question, we perform systematic numerical experiments to explore the relative computational efficiencies of several ensemble Kalman methods. The numerical experiments involve statistical observations of Lorenz‐type models of increasing complexity, frequently used to represent simplified atmospheric systems, with some featuring neural network parameterizations. For each test problem, several ensemble Kalman methods and a derivative‐based method “race” to reach a specified accuracy, and we measure the computational cost required to achieve the desired accuracy. We investigate how prior information and the parameter or data dimensions define the computational costs of the various methods.

We demonstrate a data‐driven parameterization framework for snow albedo evolution using a constrained auto‐regressive neural differential equation that directly predicts snow albedo change from standard meteorological inputs. After training with multi‐year in situ and satellite observations from a wide variety of locations, the scheme effectively reproduces daily albedo evolution across diverse climate zones, with median error under 7.5% (RMSE ≈ 0.05), a 10%–30% improvement over established models. Furthermore, the model generalizes to sites not seen during training and scales from coarser grids to point locations. The scheme can easily incorporate new features as observational networks expand, offering an adaptive and computationally lightweight framework for next‐generation land and climate models.

Uncertainty in marine low-level cloud feedbacks limits accurate climate projections. Using 7083 high-resolution simulations of tropical Pacific low clouds, we separated the impacts of sea surface warming from direct carbon dioxide (CO ₂ ) effects. Surface warming alone drives a positive low-cloud feedback of 0.14 W m − 2  K − 1 . While this changes little with doubled CO ₂ , the total cloud radiative response strengthens markedly to 0.43 W m − 2  K − 1 under quadrupled CO ₂ , revealing a strong nonlinear interaction. Surface warming alone modifies the boundary layer through increased inversion strength and weakened subsidence, which buffers clouds by promoting higher liquid water content while cloud fraction decreases. Rapid adjustments to high CO ₂ concentrations counteract this protective cloud thickening. Consequently, a pronounced reduction in cloud fraction is no longer offset by an increase in cloud brightness, markedly strengthening the total radiative response under quadrupled CO ₂ . Ultimately, our results suggest that climate sensitivity is more state-dependent than often assumed.

Quantifying the probability of rare climate extremes is essential for risk assessment, infrastructure design, and adaptation planning. Generalized extreme value (GEV) distributions provide a widely used statistical framework for estimating the return levels of low-probability, high-impact events, but such estimates are often derived from limited observational or simulated records. Using 13 000 years of simulations from an idealized, and thus computationally inexpensive, atmosphere-only climate model, we show that short time series ( < 30 years) generally lead to substantial underestimation of annual maximum daily precipitation and temperature return levels (through underestimation of the GEV shape parameter ξ ). The minimum length required for approximately unbiased estimation varies by variable and latitude, but generally lies between 30 and 50 years. Beyond this threshold, ensembles of multiple segments can be used to precisely and unbiasedly quantify uncertainty in rare-event statistics. Even when total data volume is large, ensembles composed of shorter segments retain substantial bias. Applying this framework to a 500 year Geophysical Fluid Dynamics Laboratory-ESM4 simulation reveals similar behavior, with the shape parameter as the key predictor of return level uncertainty and bias. After biases have been removed, further reduction of uncertainty requires data records longer than 50 years. By linking theoretical bias behavior to physically interpretable climate features, this work provides practical guidance for designing and analyzing datasets used in climate hazard assessment. Long-duration ensemble members or observational segments are necessary for robust GEV-based estimates of low-probability, high-impact events.

Physical Review Different approaches to using data-driven methods for subgrid-scale closure modeling of geophysical turbulence have emerged recently. Most of these approaches are data hungry and lack interpretability and out-of-distribution generalizability. Here, we use a hybrid approach that combines turbulence theory, physics-based modeling, and data-driven methods to overcome these challenges. Specifically, we address the parametric uncertainty of well-known physics-based large-eddy simulation (LES) closures: the Smagorinsky (Smag) and Leith eddy-viscosity models (one free parameter) and the Jansen-Held (JH) backscattering model (two free parameters). For various cases of two-dimensional turbulence, optimal parameters are first learned online from data via ensemble Kalman inversion (EKI), such that for each case, the LES energy spectrum matches that of direct numerical simulation (DNS). We quantify the uncertainties on these parameters using a modern machine-learning-accelerated Bayesian workflow, “, , ” Only a small training dataset is needed (to calculate the DNS spectra); i.e., the approach is data-efficient. We find the optimized parameter(s) and their associated uncertainty for each closure to be constant across broad flow regimes that differ in dominant length scales, eddy/jet structures, and dynamics, suggesting that these closures are generalizable. Next, we show that the online learned constants agree with the predictions of a recent semianalytical derivation, providing further interpretability. In both and tests that include examining the extreme events, LES with optimized closures, especially with JH, outperforms the baselines (LES with standard Smag, dynamic Smag, or Leith). This work shows the promise of combining advances in theory, physics-based modeling (e.g., JH), and data-driven modeling (e.g., online learning with EKI) to develop data-efficient frameworks for accurate, interpretable, and generalizable closures for geophysical turbulence, with ultimate applications in weather and climate prediction.

The editors of Journal of Advances in Modeling Earth Systems thank the 1,035 reviewers who provided 1,649 reviews during 2025. Their hard work and insights, typically done anonymously, benefits authors, readers, and the broader science community.

Are AI-powered climate models intrinsically more efficient than traditional climate models?While progress is still needed before they become operational, hybrid AI-physics climate models and AI emulators of climate models have the potential to sharply reduce inference cost relative to traditional CPU-based models, allowing larger ensembles to explore different scenarios and sharpen uncertainty estimation. Yet this apparent efficiency becomes less obvious when the comparison includes GPU-ported dynamical climate models, and when efficiency is assessed against the effective complexity of the simulated climate system.As a first step, recognizing that a perfect apple-to-apple comparison is rarely possible from reported configurations, we synthesize reported performance for leading AI climate model emulators (e.g., ACE2, CAMulator), hybrid AI-physics models (e.g., CliMA, NeuralGCM), and GPU-accelerated traditional models (e.g., SCREAM, ICON). We examine two complementary scaling views. The first compares throughput (simulated years per day) per accelerator (GPUs or TPUs) and per prognostic variable, as a function of horizontal grid spacing. The second compares the same normalized throughput against an effective complexity proxy, defined as the number of vertical levels divided by the product of the time step and the squared horizontal grid spacing, to account for the simulated vertical structure and, importantly, time-step constraints imposed by numerical stability.We find that AI-powered models can show favorable apparent scaling with horizontal resolution in raw throughput, but that the advantage becomes modest once effective complexity is accounted for: at comparable complexity, AI climate models do not appear intrinsically more efficient than GPU-ported dynamical models. Hybrid approaches occupy a distinct middle ground: their acceleration and added value come primarily from learned parameterizations that improve the representation of unresolved processes while the overall model retains a physically-based dynamical core, including explicit conservation laws. AI climate model emulators, by contrast, offer their clearest computational advantage through task-targeted prediction, where a limited set of climate-relevant variables can be directly simulated on the grid of interest. This avoids integrating the full high-frequency, multivariate state at the short time step traditionally required for numerical stability, which is especially advantageous when emulating a fine-resolution reference model with a coarser emulator. Diverse downscaling or targeted post-processing strategies can further substitute for explicit fine-scale resolution when observations are available, enabling inexpensive local or hazard-specific risk assessment at decadal to multi-decadal time horizons.