Advanced Climate Modeling

Further progress in AI+climate modeling "Applying the ACE2 Emulator to SST Green's Functions for the E3SMv3 Global Atmosphere Model". Building on ACE2 model which uses our spherical Fourier neural operator (SFNO) architecture, this work shows that ACE2 can replicate climate model responses to sea surface temperature perturbations with high fidelity at a fraction of the cost. This accelerates climate sensitivity research and helps us better understand radiative feedbacks in the Earth system. Background: The SFNO architecture was first used in training FourCastNet weather model, whose latest version (v3) has state-of-art probabilistic calibration. AI+Science is not just about blindly applying the standard transformer/CNN "hammer". It is about carefully designing neural architectures that incorporate domain constraints like geometry and multiple scales, while being expressive and easy to train. SFNO accomplishes both: it incorporates multiple scales, and it respects the spherical geometry and this is critical for success in climate modeling. Unlike short-term weather, which requires only a few autoregressive steps for rollout, climate modeling requires long rollouts with thousands or even greater number of time steps. All other AI-based models fail for long-term climate modeling including Pangu and GraphCast which ignore the spherical geometry. Distortions start building up at the poles since the models assume domain is a rectangle, and they lead to catastrophic failures. Structure matters in AI+Science!

Precipitation is one of the most challenging variables to accurately simulate in global climate models as it depends on small-scale physical processes. In our latest research published in 𝘚𝘤𝘪𝘦𝘯𝘤𝘦 𝘈𝘥𝘷𝘢𝘯𝘤𝘦𝘴, we describe an advancement in our hybrid atmospheric model, NeuralGCM, which now leverages AI trained directly on NASA satellite observations to improve global precipitation simulations. Key results of this work: 👉 Physics-AI Integration: The model combines a traditional fluid dynamics solver for large-scale processes with AI neural networks that learn to account for the effects of small-scale physics, specifically precipitation. 👉 Improved Extremes: NeuralGCM demonstrates significant improvements in capturing the intensity of the top 0.1% of extreme rainfall events, better representing heavy precipitation than many traditional models. 👉 Long-Term Accuracy: In multi-year simulations, the model achieved a 40% average error reduction over land compared to leading atmospheric models used in the latest Intergovernmental Panel on Climate Change (IPCC) report. 👉 Daily Patterns: It more accurately reproduces the timing of peak daily precipitation, which is critical for hydrology and agricultural planning. We are already seeing the value of this approach in the field. A partnership between the University of Chicago and the Indian Ministry of Agriculture recently used NeuralGCM in a pilot program to help predict the onset of the monsoon season. NeuralGCM is part of our Earth AI program to better understand the physical earth in ways that benefit society. We have made the code and model checkpoints openly available to the community. Read the full details on the Google Research blog by Janni Yuval: goo.gle/4qH63sU Paper: https://lnkd.in/d7E4US4W

AI finds a missing equation for simulating atmospheric and oceanic turbulence Climate models and weather forecasts simulate turbulent flows spanning scales from thousands of kilometers down to meters. No computer can resolve all of them, so modelers approximate the effect of unresolved small scales on the large-scale dynamics. This approximation is known as a subgrid-scale closure—a model that "closes" the governing equations by filling in what the coarse grid cannot see. Getting closures right matters enormously: their shortcomings are a leading source of uncertainty in climate projections and extreme weather forecasts. For decades, the field has faced a trade-off. One family of closures faithfully reconstructs the small-scale stress patterns but makes simulations blow up. The other keeps simulations stable but oversimplifies the physics—removing too much energy, ignoring backscatter from small to large scales, and underestimating extreme events. Karan Jakhar, Yifei Guan, and Pedram Hassanzadeh break this impasse by changing what equation discovery optimizes for. Previous sparse regression searches consistently landed on the same second-order approximation known since the 1970s, which is accurate but unstable. The key insight: if you also require the discovered equation to reproduce how energy flows between scales—not just match local stress patterns—the algorithm finds something different. Searching 930 candidate terms with this physics-informed dual criterion, Bayesian sparse regression robustly identifies an additional fourth-order term in the Taylor expansion of the subgrid stress (NGM4). NGM4 achieves ~0.99 pattern correlation with reference data, produces stable simulations across four diverse 2D turbulence setups mimicking atmospheric and oceanic dynamics, and accurately captures both bulk statistics and rare extremes. Its coefficients depend only on grid resolution—no tuning for flow regime or Reynolds number—and it needs just 100 training snapshots. The most striking aspect: NGM4 could have been derived analytically decades ago, but because the source of the second-order instability was unclear, higher-order terms were never explored. It took sparse regression guided by the right physics to reveal that the missing piece had been hiding in plain sight. One takeaway that extends well beyond turbulence: the criterion you optimize for determines what you discover. Embedding the right physics into equation discovery can uncover interpretable, generalizable equations that purely data-driven approaches systematically miss. Paper: https://lnkd.in/exGQGaGc #MachineLearning #Turbulence #ClimateModeling #EquationDiscovery #AIforScience #LargeEddySimulation #GeophysicalFluidDynamics #SparseRegression #PhysicsInformedAI #SubgridModeling #DeepLearning #ComputationalPhysics #ExtremeEvents #WeatherPrediction #AIforClimate

You might have seen news from our Google DeepMind colleagues lately on GenCast, which is changing the game of weather forecasting by building state-of-the-art weather models using AI. Some of our teams started to wonder – can we apply similar techniques to the notoriously compute-intensive challenge of climate modeling? General circulation models (GCMs) are a critical part of climate modeling, focused on the physical aspects of the climate system, such as temperature, pressure, wind, and ocean currents. Traditional GCMs, while powerful, can struggle with precipitation – and our teams wanted to see if AI could help. Our team released a paper and data on our AI-based GCM, building on our Nature paper from last year - specifically, now predicting precipitation with greater accuracy than prior state of the art. The new paper on NeuralGCM introduces 𝗺𝗼𝗱𝗲𝗹𝘀 𝘁𝗵𝗮𝘁 𝗹𝗲𝗮𝗿𝗻 𝗳𝗿𝗼𝗺 𝘀𝗮𝘁𝗲𝗹𝗹𝗶𝘁𝗲 𝗱𝗮𝘁𝗮 𝘁𝗼 𝗽𝗿𝗼𝗱𝘂𝗰𝗲 𝗺𝗼𝗿𝗲 𝗿𝗲𝗮𝗹𝗶𝘀𝘁𝗶𝗰 𝗿𝗮𝗶𝗻 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻𝘀. Kudos to Janni Yuval, Ian Langmore, Dmitrii Kochkov, and Stephan Hoyer! Here's why this is a big deal: 𝗟𝗲𝘀𝘀 𝗕𝗶𝗮𝘀, 𝗠𝗼𝗿𝗲 𝗔𝗰𝗰𝘂𝗿𝗮𝗰𝘆: These new models have less bias, meaning they align more closely with actual observations – and we see this both for forecasts up to 15 days, and also for 20-year projections (in which sea surface temperatures and sea ice were fixed at historical values, since we don’t yet have an ocean model). NeuralGCM forecasts are especially performant around extremes, which are especially important in understanding climate anomalies, and can predict rain patterns throughout the day with better precision. 𝗖𝗼𝗺𝗯𝗶𝗻𝗶𝗻𝗴 𝗔𝗜, 𝗦𝗮𝘁𝗲𝗹𝗹𝗶𝘁𝗲 𝗜𝗺𝗮𝗴𝗲𝗿𝘆, 𝗮𝗻𝗱 𝗣𝗵𝘆𝘀𝗶𝗰𝘀: The model combines a learned physics model with a dynamic differentiable core to leverage both physics and AI methods, with the model trained directly on satellite-based precipitation observations. 𝗢𝗽𝗲𝗻 𝗔𝗰𝗰𝗲𝘀𝘀 𝗳𝗼𝗿 𝗘𝘃𝗲𝗿𝘆𝗼𝗻𝗲! This is perhaps the most exciting news! The team has made their pre-trained NeuralGCM model checkpoints (including their awesome new precipitation models) available under a CC BY-SA 4.0 license. Anyone can use and build upon this cutting-edge technology! https://lnkd.in/gfmAx_Ju 𝗪𝗵𝘆 𝗧𝗵𝗶𝘀 𝗠𝗮𝘁𝘁𝗲𝗿𝘀: Accurate predictions of precipitation are crucial for everything from water resource management and flood mitigation to understanding the impacts of climate change on agriculture and ecosystems. Check out the paper to learn more:  https://lnkd.in/geqaNTRP

🌍 Climate scientists often face a trade-off: Global Climate Models (GCMs) are essential for long-term climate projections — but they operate at coarse spatial resolution, making them too crude for regional or local decision-making. To get fine-scale data, researchers use Regional Climate Models (RCMs). These add crucial spatial detail, but come at a very high computational cost, often requiring supercomputers to run for months. ➡️ A new paper introduces EnScale — a machine learning framework that offers an efficient and accurate alternative to running full RCM simulations. Instead of solving the complex physics from scratch, EnScale "learns" the relationship between GCMs and RCMs by training on existing paired datasets. It then generates high-resolution, realistic, and diverse regional climate fields directly from GCM inputs. What makes EnScale stand out? ✅ It uses a generative ML model trained with a statistically principled loss (energy score), enabling probabilistic outputs that reflect natural variability and uncertainty ✅ It is multivariate – it learns to generate temperature, precipitation, radiation, and wind jointly, preserving spatial and cross-variable coherence ✅ It is computationally lightweight – training and inference are up to 10–20× faster than state-of-the-art generative approaches ✅ It includes an extension (EnScale-t) for generating temporally consistent time series – a must for studying events like heatwaves or prolonged droughts This approach opens the door to faster, more flexible generation of regional climate scenarios, essential for risk assessment, infrastructure planning, and climate adaptation — especially where computational resources are limited. 📄 Read the full paper: EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules ---> https://lnkd.in/dQr5rmWU (code: https://lnkd.in/dQk_Jv8g) 👏 Congrats to the authors — a strong step forward for ML-based climate modeling! #climateAI #downscaling #generativeAI #machinelearning #climatescience #EnScale #RCM #GCM #ETHZurich #climatescenarios

Climate models have long struggled with coarse resolution, limiting precise climate risk insights. But AI-driven methods are now changing this, unlocking more detailed intelligence than traditional physics-based approaches. I recently spoke with a research scientist at Google Research who highlighted a promising new hybrid approach. This method combines physics-based General Circulation Models (GCMs) with AI refinement, significantly improving resolution. The process starts with Regional Climate Models (RCMs) anchoring physical consistency at ~45 km resolution. Then, it uses a diffusion model, R2-D2, to enhance output resolution to 9 km, making estimates more suitable for projecting extreme climate events. 🔥 About R2-D2 R2‑D2 (Regional Residual Diffusion-based Downscaling) is a diffusion model trained on residuals between RCM outputs and high-resolution targets. Conditioned on physical inputs like coarse climate fields and terrain, it rapidly generates high-res climate maps (~800 fields/hour on GPUs), complete with uncertainty estimates. ✅ Why this matters - Offers detailed projections of extreme climate events for precise risk quantification. - Delivers probabilistic forecasts, improving risk modeling and scenario planning. - Provides another high-resolution modeling approach, enriching ensemble strategies for climate risk projections. 👉 Read the full paper: https://lnkd.in/gU6qmZTR 👉 An excellent explainer blog: https://lnkd.in/gAEJFEV2 If your work involves climate risk assessment, adaptation planning, or quantitative modeling, how are you leveraging high-resolution risk projections?

🔥 𝗡𝗲𝘄 𝗽𝗮𝗽𝗲𝗿 𝗶𝗻 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 𝗔𝗱𝘃𝗮𝗻𝗰𝗲𝘀 Permafrost may become a net source of CO₂ much earlier than previously projected. Our new study shows that Earth System Models have been missing a major piece of the puzzle: deep frozen carbon stored in Yedoma deposits and northern peatlands. Using an enhanced version of the ORCHIDEE-MICT model that explicitly represents the formation and evolution of these deep carbon pools over thousands of years, we find that: 🔹 Deep permafrost carbon is far more vulnerable to thaw than previously assumed. 🔹 Including these carbon reservoirs reduces the projected cumulative northern soil carbon accumulation by 47–74 PgC by 2100. 🔹 Under high-emission scenarios, northern soils will shift from a net carbon sink to a net carbon source during this century, releasing up to 32 PgC. 🔹 The transition from carbon uptake to carbon loss occurs decades earlier than projected by many current Earth System Models. These results suggest that the permafrost carbon-climate feedback may be stronger and faster than currently represented in CMIP-class models, highlighting the urgent need to improve the representation of deep carbon processes in climate projections. 📄 Paper: https://lnkd.in/ejJnewwH I am delighted to see this work led by Yi Xi at Laboratoire des Sciences du Climat et de l'Environnement and colleagues now published in Science Advances. This research benefited from developments by the CALIPSO project led by UVSQ Université de Versailles Saint-Quentin-en-Yvelines CEA CNRS and funded by Schmidt Sciences, which aims to advance next-generation Earth system modelling by better representing critical climate feedbacks, including those associated with permafrost carbon. #ClimateChange #Permafrost #CarbonCycle #EarthSystemScience #Arctic #ClimateModeling #NetZero #ScienceAdvances #CALIPSO #SchmidtSciences

I am very excited about our new study that was just published in Nature Geoscience. It shows that future extreme precipitation will intensify far more than previously estimated, driven by stronger mesoscale moisture convergence. Using a global high-resolution Earth system model that simulates extreme-producing phenomena far better than its low-resolution counterpart (see image below), we find that daily extremes could rise by over 40% by 2100—nearly three times the dynamical contribution seen in standard low-resolution models. These results highlight the urgent need for high-resolution climate modeling to constrain risks better and support effective adaptation strategies.

Excited to share new NVIDIA Earth-2 research on km-scale global climate foundation modeling with generative AI. "Climate In a Bottle" avoids autoregression and leverages conditional diffusion for flexibility, multi-modal generation, and cascaded super-resolution.    Preprint: https://lnkd.in/ggYWnse9 Code: https://lnkd.in/grCEMrxb   I find it frankly mind-boggling to be able to synthesize a dozen ~13M-pixel atmospheric channels at planetary scales from only 200 kB of climate-controlling input boundary conditions. And mixing information from ERA5 reanalysis vs. ICON simulation modalities is very fun.   Acknowledging a fantastic team - Noah Brenowitz, Tao Ge, Akshay Subramaniam, Aayush Gupta, David Hall, Morteza Mardani, Arash Vahdat and Karthik Kashinath.

Excited to share that our manuscript, “Physics-informed deep learning reveals climate-driven snowpack decline and threatens ecological water availability in a Californian snow-fed catchment”, led by my PhD student Surendra Maharjan has been published in Ecological Informatics (Impact Factor – 7.3, CiteScore – 11.4). This study focuses on the Upper West Walker River Watershed in the eastern Sierra Nevada, California, a mountainous snow-fed region where seasonal snowpack acts as a natural reservoir, storing water in winter and releasing it gradually to sustain streamflow and ecological systems downstream. However, climate warming is shifting precipitation from snow to rain and accelerating melt, increasing vulnerability and creating an urgent need for advanced tools that can forecast ecological water risks. To address this challenge, this study evaluates three modeling approaches: the process-based SWAT hydrological model, a data-driven Long Short-Term Memory (LSTM) deep learning model, and a Physics-Informed LSTM (PIML) that integrates melt physics and precipitation-phase constraints. Key Findings: The PIML model demonstrated the most robust and well-balanced performance across key hydrologic metrics (NSE, KGE, RMSE). Future climate projections indicate that peak SWE may decline by up to 60%. Peak discharge may decrease by about 33% under warming conditions. Snowmelt and runoff may shift 10–19 days earlier, shortening the hydrologic season. These changes compress the hydrologic season, threaten summer ecological water availability, and heighten drought risk across snow-fed systems. The results underscore the growing challenges of managing water resources in snow-dominated basins under climate change. Coupling physics with deep learning offers a promising path toward more reliable forecasting of snowpack dynamics and streamflow in mountain watersheds. EssDs Chapman Chapman University Schmid College of Science and Technology Surendra Maharjan Wenzhao Li Rejoice Thomas Shahryar Fazli Hesham Morgan Mohamed Allali Ali Elgendy Link : https://lnkd.in/g4f6VpiK