SamudrACE-E3SMv3

What is SamudrACE-E3SMv3?

SamudrACE-E3SMv3 is a coupled climate-model emulator provided by the Allen Institute for AI and its research collaborators. It is intended to reproduce the behavior of a preindustrial-control simulation from the Energy Exascale Earth System Model version 3, commonly called E3SMv3.

Rather than calculating every process at the same computational cost as the original Earth-system model, the emulator learns how atmospheric and oceanic state variables evolve over time. It then generates new simulated states during inference. This makes it possible to study long climate trajectories and large ensembles more quickly than with the reference simulation, while retaining important interactions between the atmosphere and ocean.

The model is not a conversational system, a text-generation model, or a general-purpose forecasting API. Its inputs and outputs are structured climate-simulation data, and practical use requires the associated checkpoint files, forcing data, initial conditions, inference configuration, and emulation software.

How the model combines atmosphere and ocean emulation

SamudrACE-E3SMv3 is built from two main components. ACE2S is the stochastic atmospheric emulator, while Samudra is the full-depth ocean emulator. The coupled system passes information between these components so that atmospheric and oceanic conditions can influence one another during a rollout.

“Stochastic” means that the atmospheric component can represent internally generated variability rather than producing only one fixed, deterministic trajectory from a given state. This matters in climate research because two simulations with the same broad external setup can still develop different weather and climate patterns through internal variability.

The atmospheric component operates at a six-hour time step, with one-degree horizontal resolution and eight vertical levels. The ocean component represents the full depth of the ocean. These are verified characteristics of the described research configuration, not general specifications for an API or cloud service.

Training and evaluation data

The model was trained on 105 years of E3SMv3 preindustrial-control data. The reported split used 90 years for training, five years for validation, and ten years for testing. The emulator was also evaluated on an independent 400-year segment of the E3SMv3 control simulation.

This evaluation design is important when interpreting the results. The model is designed around a preindustrial-control configuration, and the published work demonstrates performance for that setting. It does not establish that the released checkpoint can reliably simulate climate responses to time-varying anthropogenic radiative forcing. The research identifies that type of application as future work rather than a demonstrated capability of this release.

What the emulator can reproduce

The reported results show that SamudrACE-E3SMv3 reproduces the reference model’s mean climate state, with biases described as substantially smaller than many existing differences between climate models and observations. It also maintains long-timescale variability during extended rollouts, including important features of El Niño–Southern Oscillation behavior.

The stochastic approach improves several aspects of internal variability compared with a deterministic SamudrACE baseline. Reported areas of improvement include variability in eddy-rich sea-surface temperatures, sea ice near the marginal ice zone, and the power spectrum of ENSO-related variability.

Daily precipitation distributions are represented accurately through approximately the 99.99th percentile in the associated evaluation. However, the model underestimates the rarest tropical precipitation extremes. That distinction is significant: strong performance across most of a distribution does not make the emulator suitable by itself for analyzing the most unusual extreme events.

Speed, computational cost, and trade-offs

The associated research reports that SamudrACE-E3SMv3 can run approximately 40 times faster than the original E3SMv3 simulation on a single graphics processor. The conventional climate simulation, by contrast, requires thousands of processor cores in the referenced comparison.

This reported speed advantage is the model’s main practical reason for existing. Faster rollouts can make large ensembles more accessible, allowing researchers to examine a wider range of internally varying outcomes or run longer experiments than would be practical with the original model. The gain is not equivalent to free or turnkey computation: users still need suitable GPU hardware, the model checkpoints, climate datasets, and the accompanying inference environment.

The trade-off is that an emulator approximates the behavior of its training and evaluation system. It may be less reliable outside the represented climate configuration, and its errors can matter most for applications focused on rare extremes or previously unseen forcing conditions. A conventional Earth-system model may therefore remain more appropriate when physical process fidelity, established forcing scenarios, or extrapolation beyond the demonstrated domain is the priority.

Availability, pricing, and supported data

SamudrACE-E3SMv3 is distributed as an open research checkpoint through the Allen Institute for AI’s Hugging Face organization. The repository provides coupled, atmospheric, and ocean checkpoint archives, along with forcing data, initial conditions, and an inference configuration. The supplied research identifies the release as being available under the Apache 2.0 license.

There is no documented recurring price, token price, or managed inference endpoint for this release. “Open” refers to the research checkpoint and associated materials, not necessarily to zero total operating cost. Hardware, storage, data movement, software setup, and researcher time can still contribute substantially to deployment cost.

Inputs, outputs, and capabilities

The model works with structured atmospheric and oceanic climate-state data rather than text, images, audio, or video. Its outputs are simulated climate variables and related state fields. It therefore does not provide text output, image generation, speech, embeddings, function calling, web search, streaming responses, or a general JSON mode in the conventional API sense.

There is no documented natural-language reasoning or coding capability. Researchers may use code to operate the model and analyze its results, but that should not be confused with the model itself generating or executing software. Similarly, the repository’s inference configuration is an operational resource, not evidence of a tool-use interface.

No general context window or maximum output-token limit is specified. The relevant constraints are the supported climate-data format, model configuration, available memory, rollout setup, and the research domain represented by the checkpoint.

Best use cases

• Running large ensembles to study internal climate variability.
• Examining ENSO behavior and related long-timescale variability.
• Analyzing sea-surface temperature, precipitation, and sea-ice behavior.
• Testing coupled climate-emulation methods over long rollouts.
• Exploring computationally efficient Earth-system simulation on GPU hardware.

For example, a research group could use the emulator to generate many internally different climate trajectories and compare the resulting distributions of sea-surface temperature or precipitation. A researcher studying computational methods could instead compare stochastic and deterministic emulation behavior or evaluate how well the coupled atmosphere–ocean system preserves variability over extended simulations.

Limitations and when to choose another option

SamudrACE-E3SMv3 is a strong fit when the priority is faster experimentation with a demonstrated E3SMv3 preindustrial-control configuration. Its combination of stochastic atmospheric behavior and full-depth ocean coupling is more relevant to climate-variability research than a model that emulates only one component of the system.

Another option may be more appropriate when the project requires accurate estimates of the rarest tropical precipitation extremes, direct simulation of changing anthropogenic forcing, or a broadly validated climate scenario framework outside the published evaluation domain. The original E3SMv3 simulation may also be preferable when the additional computational expense is justified by the need to use the conventional model directly.

It is also the wrong choice for users seeking a language model, image generator, multimodal assistant, hosted prediction API, or turnkey forecasting product. SamudrACE-E3SMv3 is a specialized scientific research artifact. Its value comes from accelerating a particular class of coupled climate simulations, not from offering general-purpose AI interaction.

Overall assessment

SamudrACE-E3SMv3 occupies a specialized position in the Allen Institute for AI’s research catalog: it is an open, coupled Earth-system emulator aimed at reducing the cost of long climate simulations and large ensembles. The reported approximately 40-fold speed comparison on one GPU is useful for understanding its motivation, while the documented limitations define where caution is required.

Researchers should treat its results as emulated E3SMv3 behavior within a specified training and evaluation setting, not as a universal replacement for a conventional climate model. When used within that boundary, its stochastic atmosphere–ocean design offers a practical way to investigate climate variability at a scale that may otherwise require much larger computational resources.