In the struggle to model the Earth’s changing climate with both speed and fidelity, researchers have long been caught between scientific ambition and computational constraint. Traditional global climate simulators gloss over small-scale processes such as thunderstorms and tropical cloud systems, phenomena that are disproportionately powerful in shaping the planet’s atmospheric dynamics. The models that can capture them, called cloud-resolving models (CRMs), demand so much computing power that they remain largely impractical for long-term forecasting.
But a breakthrough in hybrid AI-physics simulation may be about to change that. NVIDIA, alongside a global consortium of researchers supported by Columbia University and the US National Science Foundation, has introduced ClimSim-Online: a reproducible, open framework that allows machine learning models to plug directly into a full-featured climate simulator.
Built on the award-winning ClimSim dataset, the framework allows AI to learn from terabytes of high-resolution simulation data and then emulate sub-grid climate physics, running tens to hundreds of times faster than traditional methods, without sacrificing accuracy or stability.
From nested physics to hybrid inference
At the heart of ClimSim-Online is the Energy Exascale Earth System Model-Multiscale Modeling Framework (E3SM-MMF), an experimental climate model that embeds thousands of CRMs inside a coarser global model. These embedded models simulate microphysical cloud dynamics at resolutions as fine as 2km. Over ten years of simulation time, the system produced nearly six billion training samples, each mapping how fine-scale processes such as convection, droplet formation, or radiation alter the large-scale atmospheric state.
This immense dataset powers machine learning models that effectively replace the embedded CRMs, which consume 95 per cent of computational resources in conventional setups. The result: dramatic reductions in cost and time, combined with the ability to run long-term simulations that maintain realism at both cloud scale and global scale.
What sets ClimSim-Online apart is its operational flexibility. NVIDIA has containerised the workflow, enabling researchers to deploy hybrid models with standardised diagnostics on workstations, HPC clusters or cloud VMs. With only a TorchScript model file and minimal dependencies, developers can integrate AI directly into the Fortran-based E3SM codebase.
New benchmarks in hybrid simulation
The release comes on the heels of a milestone paper, published 10 July in Journal of Advances in Modeling Earth Systems, where NVIDIA researchers demonstrated multi-year stable hybrid climate simulations for the first time. Using a U-Net neural network trained on ClimSim data with NVIDIA’s PhysicsNemo framework, the hybrid model stayed within 2°C temperature bias and 1g/kg humidity bias over the troposphere, an unprecedented result.
Critically, the researchers embedded hard physical constraints within the neural network itself. These include temperature-based phase partitioning for cloud condensates and an upper limit on ice clouds above the tropopause. Without such measures, previous hybrid models were prone to atmospheric drift, especially in tropical regions.
The team also achieved simulations spanning more than five years with real geographic data and fully coupled land-atmosphere dynamics. Until now, such longevity and complexity had been unattainable in AI-augmented climate systems.
Opening new doors for AI-climate collaboration
By lowering the technical barriers to hybrid modelling, ClimSim-Online invites wider collaboration between AI researchers and climate scientists. Already, the ClimSim dataset has underpinned a global Kaggle competition with more than 460 teams contributing benchmark models, many of which can now be evaluated inside real climate simulators, not just on offline metrics.
The work also opens avenues for experimentation beyond traditional scientific approaches. Reinforcement learning, for example, could offer a path to optimise hybrid models where direct gradient-based methods fail due to the non-differentiable nature of host simulators. With ClimSim-Online now making the downstream reward signals accessible, such interdisciplinary innovation may not be far away.
While challenges remain in refining hybrid models to meet policy-grade standards, this development marks a critical inflection point. The combination of physics-based reasoning and machine learning acceleration is no longer aspirational, it is operational, and increasingly, indispensable.
In an age of climate volatility, the ability to simulate the planet’s atmosphere with both nuance and speed may be one of AI’s most consequential applications yet.
