The AIGS Pod

The AIGS

5.0 (1)
Natural Sciences
Updated Biweekly

The AIGS podcast tackling AI in earth system modeling and more.

3d ago

Why climate AI needs synthetic data

As deep learning for physical systems continues to grow in popularity, efforts to improve generalizability have primarily focused on designing architectures that embed physical constraints. However, for machine-learning surrogate climate models (emulators), research shows that the low structural diversity in existing scenarios commonly used to generate training data places a ceiling on predictive skill. This work examines whether training datasets themselves can be optimized to improve generalization. The authors introduce a method to create datasets that produce emulators capable of generalizing to new, structurally different scenarios absent from the training data. They use a differentiable Simple Climate Model (SCM) to calculate the sensitivity of emulator loss to perturbations in the training data, iteratively updating the training data to maximize emulator skill. Training on one scenario optimized in this fashion outperforms an emulator trained on six standard ScenarioMIP pathways while using a smaller dataset. The emulator successfully isolates distinct physical behaviors of different climate forcing agents (e.g., greenhouse gases vs. aerosols) without single-forcing runs. Results suggest that generating a small number of dynamically rich scenarios provides greater marginal value for emulation and characterizing system responses than expanding the suite of traditional emissions pathways. Paper: https://arxiv.org/abs/2606.19302v1
Jul 6

The Tebaldi Test For Climate Emulators

A comprehensive review of climate model emulation methods and their validation. This episode discusses the latest survey from Claudia Tebaldi, Noelle E. Selin, Rafael Ferrari, and Glenn Flierl, published in the Annual Review of Environment and Resources (2025). The paper surveys modern approaches to climate model emulation—surrogate models that replicate Earth System Model behavior at a fraction of the computational cost. Key topics include: statistical and machine learning emulation methods; trade-offs between model simplicity and accuracy; handling climate system internal variability; applications to anthropogenic climate forcing scenarios; and validation approaches for high-dimensional climate outputs. The Tebaldi Test represents a critical benchmark for evaluating whether emulators can generalize beyond their training data and maintain physical consistency under extreme climate scenarios. Paper: https://www.annualreviews.org/content/journals/10.1146/annurev-environ-012125-085838
Jun 30

Solving climate emulator physics with random CO2

Machine-learning weather-climate emulators like the Ai2 Climate Emulator (ACE) can reproduce the climate of recent decades, but they break down in unfamiliar conditions, producing unphysical results when asked to simulate a world with sea surface temperatures warmed by +4 K or an abrupt quadrupling of CO2. This research traces the problem to training data in which sea surface temperature (SST) and CO2 are tightly correlated, leaving the models unable to learn their separate effects. To fix this, the authors introduce a new class of "random-CO2" reference simulations: deliberately stitched-together climates in which SST and CO2 are prescribed to vary independently. By training on a balance of AMIP, equilibrium-climate, and random-CO2 data, and adding a total energy conservation constraint for improved interpretability, they build a more data-efficient emulator that not only matches earlier models where they already excelled, but also handles previously failing scenarios like AMIP +4 K and slab-ocean-coupled abrupt 4xCO2. The model still relies on simplified or prescribed ocean, land, and sea-ice components and inherits biases from its physics-based training data, each pointing to clear avenues for future work. Paper: https://arxiv.org/abs/2606.07928
Jun 30

AI learns physics from Frankenstein climates

Machine-learning weather-climate emulators like the Ai2 Climate Emulator (ACE) can reproduce the climate of recent decades, but they break down in unfamiliar conditions, producing unphysical results when asked to simulate a world with sea surface temperatures warmed by +4 K or an abrupt quadrupling of CO2. This research traces the problem to training data in which sea surface temperature (SST) and CO2 are tightly correlated, leaving the models unable to learn their separate effects. To fix this, the authors introduce a new class of "random-CO2" reference simulations: deliberately stitched-together "Frankenstein" climates in which SST and CO2 are prescribed to vary independently. By training on a balance of AMIP, equilibrium-climate, and random-CO2 data, and adding a total energy conservation constraint for improved interpretability, they build a more data-efficient emulator that not only matches earlier models where they already excelled, but also handles previously failing scenarios like AMIP +4 K and slab-ocean-coupled abrupt 4xCO2. The model still relies on simplified or prescribed ocean, land, and sea-ice components and inherits biases from its physics-based training data, each pointing to clear avenues for future work. Paper: https://arxiv.org/abs/2606.07928
May 26

Why AI weather models fail climate emulation

A look at the growing body of work questioning where AI weather and climate models fall short, and why physics-based simulation remains essential. Papers discussed: Zhang et al. (2026), Physics-based models outperform AI weather forecasts of record-breaking extremes, Science Advances, https://www.science.org/doi/full/10.1126/sciadv.aec1433 — Smith & Thorpe (2026), The Primacy of Physical Simulation in the Age of AI: A Critique of ML for Weather Forecasting, Bulletin of the American Meteorological Society, https://journals.ametsoc.org/view/journals/bams/aop/BAMS-D-25-0214.1/BAMS-D-25-0214.1.xml — Scaife (2026), Successes and failures of current AI climate models, Geophysical Research Letters, https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2026GL122615 — Shaw & Stevens (2025), The other climate crisis, Nature, https://www.nature.com/articles/s41586-025-08680-1
Apr 28

AI fixes climate model blind spots

Climate simulations, at all grid resolutions, rely on approximations that encapsulate the forcing due to unresolved processes on resolved variables, known as parameterizations. Parameterizations often lead to inaccuracies in climate models, with significant biases in the physics of key climate phenomena. Advances in artificial intelligence (AI) are now directly enabling the learning of unresolved processes from data to improve the physics of climate simulations. This research introduces a flexible framework for developing and implementing physics- and scale-aware machine learning parameterizations within climate models, focusing on the ocean and sea-ice components of a state-of-the-art climate model by implementing a spectrum of data-driven parameterizations, ranging from complex deep learning models to more interpretable equation-based models. The results showcase the viability of AI-driven parameterizations in operational models, advancing the capabilities of a new generation of hybrid simulations, and include prototypes of fully coupled atmosphere-ocean-sea-ice hybrid simulations. The tools developed are open source, accessible, and available to all. Paper: https://arxiv.org/abs/2510.22676
04/05/2025

AI fixes systematic climate model bias

Coarse resolution, imperfect parameterizations, and uncertain initial states and forcings limit Earth-system model (ESM) predictions. Traditional bias correction via data assimilation improves constrained simulations but offers limited benefit once models run freely. This research introduces an operator-learning framework that maps instantaneous model states to bias-correction tendencies and applies them online during integration. Building on a U-Net backbone, two operator architectures—Inception U-Net (IUNet) and a multi-scale network (M&M)—combine diverse upsampling and receptive fields to capture multiscale nonlinear features under Energy Exascale Earth System Model (E3SM) runtime constraints. Trained on two years of E3SM simulations nudged toward ERA5 reanalysis, the operators generalize across height levels and seasons. Both architectures outperform standard U-Net baselines in offline tests, indicating that functional richness rather than parameter count drives performance. In online hybrid E3SM runs, M&M delivers the most consistent bias reductions across variables and vertical levels. The ML-augmented configurations remain stable and computationally feasible in multi-year simulations, providing a practical pathway for scalable hybrid modeling. Paper: https://arxiv.org/abs/2512.03309v1
03/31/2025

Why Tunnel Vision Fixes Climate AI

This research introduces a modified machine-learning (ML) weather emulator designed to accurately predict fast radiative feedbacks in response to varying CO2 levels. While traditional emulators often struggle with global perturbations, the authors developed a column-local architecture for the Allen Institute for Artificial Intelligence Climate Emulator (ACE) to better represent atmospheric physics. By coupling this ML model with a physics-based radiative transfer scheme (RRTMG), the researchers successfully replicated the hydrological and thermal responses found in complex Earth System Models (ESMs). The study demonstrates that emulators trained only on historical climate data can still simulate unprecedented greenhouse gas scenarios by focusing on rapid atmospheric processes. These findings suggest that hybrid ML-physics models can significantly reduce the computational cost of climate projections while maintaining physical reliability. Consequently, this framework offers a powerful new tool for sampling internal atmospheric variability and conducting extensive climate sensitivity experiments. Paper: https://arxiv.org/abs/2602.16090
03/29/2025

AI weather models are stuck in 1998

This research investigates a significant cold bias in modern AI weather and climate models, such as FourCastNet, Pangu, and ACE2, which stems from their reliance on historical training data. By evaluating these models on recent time periods outside of their training sets, the authors discovered that the predicted temperatures often reflect climatic conditions from 15 to 30 years ago rather than current warming trends. The study highlights a "pull" toward the past: weather models struggle to predict extreme heat events due to a lack of modern examples, while the climate model shows the greatest inaccuracies in regions where global warming has been most rapid. Ultimately, the paper argues that even with the inclusion of CO2 data, these data-driven models remain anchored to their training-set history, necessitating new strategies to ensure they can accurately forecast an increasingly hot and unprecedented future. Paper: https://doi.org/10.1029/2025GL119740

9 Episodes

The AIGS podcast tackling AI in earth system modeling and more.

Creator

The AIGS
Years Active

2026
Episodes

9
Rating

Clean
Show Website

The AIGS Pod