The climate modeling community stands at a pivotal moment. After decades of incremental improvements, we've reached a fundamental shift in how we simulate Earth's climate system—one that promises unprecedented accuracy but demands extraordinary computational resources. Recent breakthroughs in kilometer-scale climate modeling are transforming our understanding of climate systems while creating computational challenges that are driving convergence between traditional physics-based and AI-accelerated approaches.
As climate impacts intensify globally, the World Meteorological Organization projects that advances in computing power will enable century-long global simulations with kilometer-scale models within just five years. This creates an urgent window for strategic investment in research teams pioneering scalable approaches to these computational challenges, with implications for both scientific advancement and practical climate adaptation planning.
The Resolution Revolution in Climate Modeling
Traditional climate models operate at resolutions between 50-100 kilometers, forcing them to approximate small-scale processes like cloud formation and convection through parameterizations—mathematical shortcuts that introduce significant uncertainties. Kilometer-scale models fundamentally change this paradigm by directly simulating these processes, revealing critical feedback loops between atmosphere, ocean, and land systems previously invisible in traditional models.
The European nextGEMS project exemplifies this revolution, successfully producing multi-decadal climate simulations with kilometer-scale processes across ocean, land, and atmosphere. Using the Levante supercomputer, the project has made significant advances in computational throughput, enabling multi-decadal climate simulations with kilometer-scale processes across ocean, land, and atmosphere. These models can directly simulate small-scale processes like convective cloud dynamics, reducing uncertainties from parameterizations used in traditional models.
What makes this resolution revolution particularly powerful is how it reveals connections across traditionally separate modeling domains. When atmospheric convection, ocean eddies, and land surface processes are all simulated at kilometer scales, we begin to see how these systems interact in ways that coarser models simply cannot capture. Recent extreme weather events, including the 2021 Pacific Northwest heat wave and Hurricane Otis in 2023, highlight the urgent need for enhanced local climate predictions that only kilometer-scale models can provide.
The integration becomes especially critical when examining ocean-atmosphere interactions. A global coupled simulation conducted with kilometer-scale resolution (7 km for atmosphere, 2-4 km for ocean) revealed how mesoscale sea surface temperature anomalies influence atmospheric storm tracks and ocean eddies—interactions that traditional models miss entirely. This simulation produced nearly two petabytes of model output over 14 months, demonstrating both the scientific value and data management challenges of high-resolution modeling.
The Computational Challenge of Unprecedented Resolution
This unprecedented resolution comes with extraordinary computational costs. A near-global climate simulation at 1 km resolution conducted on the Piz Daint supercomputer consumed 596 MWh per simulated year and achieved a throughput of just 0.043 simulated years per day using 4888 GPUs. To put this in perspective, traditional climate models can consume up to 10 megawatt hours to simulate an entire century of climate.
The data storage requirements are equally daunting. One week of simulated data at kilometer-scale resolution requires approximately 200 terabytes of storage, creating significant data management challenges for research institutions. The Max Planck Institute's fully coupled global atmosphere-ocean model simulation at 1 km resolution represents a significant milestone, but the DKRZ supercomputer Levante is currently the only facility capable of handling such extensive simulations.
This creates a paradox illustrated by the exponential scaling relationship: as resolution increases from 100 km to 1 km, computational requirements increase dramatically, with near-global simulations at 1 km resolution requiring substantially greater resources than traditional models. Our most realistic climate simulations may be our least sustainable to run. Most Earth System Models currently utilize less than 5% of CPU capabilities, indicating significant room for optimization, but even with perfect efficiency, the computational demands of kilometer-scale models would remain extraordinary.
Convergent Pathways in Traditional and AI-Accelerated Approaches
Research teams are pursuing parallel paths to address these computational challenges: optimizing traditional physics-based models for exascale computing while developing AI-based emulators that can achieve similar accuracy with dramatically lower energy requirements.
On the traditional optimization front, the U.S. Department of Energy's Energy Exascale Earth System Model (E3SM) has integrated super-parameterization techniques—embedding cloud-resolving models within coarse-grid climate models—to enhance cloud resolution while reducing computational time and costs. The multiscale modeling framework allows E3SM to run at higher resolutions (1-3 km) while maintaining global simulations at coarser resolutions (25-100 km). However, these simulations are still approximately 100 times more computationally expensive than traditional methods.
Meanwhile, AI-based approaches are showing remarkable efficiency gains. NVIDIA's cBottle model simulates global climate at kilometer-scale resolution and achieves data compression of up to 3,000x for individual weather samples, significantly reducing storage needs. The model was tested at the World Climate Research Programme Global KM-Scale Hackathon, indicating collaborative validation efforts across the research community.
Even more dramatically, the Ai2 Climate Emulator accelerates climate simulations by 1000x and reduces power consumption by 10,000x compared to traditional models. Training on 100 years of NOAA model data takes just 2.5 days on four NVIDIA A100 Tensor Core GPUs, and a 100-year simulation runs in three hours on a single A100 GPU.
These approaches aren't competing alternatives but complementary pathways. Traditional physics-based models provide the training data and validation benchmarks for AI emulators, while AI approaches can accelerate exploration of climate scenarios and provide rapid feedback for traditional model development.
Prediction Breakthroughs in Extreme Weather Accuracy
The ultimate test of any climate model is its prediction accuracy, and kilometer-scale models are demonstrating significant improvements, particularly for tropical cyclones and regional extreme weather events that reveal cross-domain interactions between atmospheric dynamics, ocean heat transport, and land surface processes.
Google DeepMind's Weather Lab features an AI-based tropical cyclone model that predicts cyclone formation, track, intensity, size, and shape. The model generates 50 possible scenarios up to 15 days ahead, showing improved accuracy over traditional physics-based methods. Internal evaluations indicate that the AI model's 5-day track prediction averages 140 km closer to the true cyclone location than leading global models.
A study published in npj Climate and Atmospheric Science introduced a fast, physics-based perturbation generator for AI weather models, enhancing ensemble forecasts of tropical cyclone tracks. The AI-based model significantly outperforms traditional European Centre for Medium-Range Weather Forecasts models in both deterministic and probabilistic metrics. The ensemble forecasts utilize 2000 members—a first in the field—leading to improved forecast skills for extreme scenarios of tropical cyclone movement.
Beyond tropical cyclones, kilometer-scale models reveal critical interactions between atmospheric extremes, ocean heat transport, and land surface processes that traditional models miss. When simulated at resolutions below 10 km, these cross-domain interactions significantly improve predictions of local climate extremes compared to traditional 50-100 km models, reducing uncertainties in regional climate predictions and improving accuracy for extreme weather events.
Strategic Investment Landscape for Climate Modeling Teams
For investors seeking to support the next generation of climate modeling capabilities, the landscape offers several strategic pathways with complementary risk-reward profiles. Investment-ready teams demonstrate three key characteristics: cross-disciplinary expertise spanning traditional physics and AI approaches; established partnerships with supercomputing facilities; and proven scalability from regional to global applications.
The nextGEMS project represents a significant European investment in kilometer-scale climate modeling, involving collaboration among 26 institutes across 14 European nations. The project is structured around four cycles, each releasing new configurations of models evaluated at hackathons involving diverse expertise. Models used include the ICOsahedral Non-hydrostatic model (ICON) and the Integrated Forecasting System coupled to the Finite-volumE Sea ice-Ocean Model (IFS-FESOM).
The ESCAPE-2 project focuses on energy-efficient scalable algorithms for weather and climate prediction at the exascale, involving collaboration among 12 European partners, including national meteorological services and universities. New algorithmic developments include a discontinuous Galerkin dynamical core option, enhancing computational efficiency.
In the private sector, Google DeepMind and NVIDIA are leading investments in AI-enhanced climate modeling. NVIDIA's Earth-2 platform integrates AI, GPU acceleration, and physical simulations to enhance climate predictions while offering tools for interactive visualization—a critical capability for translating complex model outputs into actionable insights.
Bridging the Computational Divide
The convergence of traditional physics-based modeling with AI-accelerated approaches represents not just a technical evolution but a fundamental paradigm shift in climate science. As these complementary approaches mature, they promise to democratize access to high-resolution climate projections while revealing previously invisible interactions between atmosphere, ocean, and land systems.
For climate scientists, this convergence demands new cross-domain collaborations that transcend traditional disciplinary boundaries. The most promising research teams are already building these bridges, integrating expertise from atmospheric physics, oceanography, computer science, and machine learning. The World Meteorological Organization's projection of century-long global simulations with kilometer-scale models within five years creates a clear timeline for deployment, with significant implications for climate adaptation planning and risk assessment.
The computational revolution in climate modeling isn't just about faster computers or more accurate predictions—it's about fundamentally transforming how we understand the interconnected systems that drive our planet's climate. By supporting this revolution, we can accelerate our understanding of regional climate impacts while reducing the environmental footprint of climate modeling itself, creating a sustainable pathway toward the climate predictions our changing world desperately needs.
Things to follow up on...
-
FengWu-GHR breakthrough model: This AI weather model operates at 0.09 km resolution and demonstrates superior forecasting skills compared to traditional numerical models, particularly for extreme weather events like heat waves and floods.
-
Generative data assimilation: Researchers have developed methods that utilize sparse weather station observations to enhance kilometer-scale weather models through score-based data assimilation, achieving 10% lower RMSEs compared to baseline systems.
-
Climate visualization challenges: Current climate data visualizations often cater to expert audiences, and effective visualizations should balance information density, robustness, and saliency to improve accessibility for broader stakeholder groups.
-
CMIP7 evaluation framework: The Coupled Model Intercomparison Project is developing systematic model evaluation tools and new diagnostics to assess the capabilities and limitations of higher resolution and complexity models participating in CMIP7.

