New applications of the Ai2 Climate Emulator (ACE) by the international climate modeling community

Ai2’s Climate Modeling team launched Version 2 of the Ai2 Climate Emulator (ACE2) in 2024 (Watt-Meyer et al. 2024, 2025). This included a model trained on ERA5, an observational estimate of global weather every 6 hours from 1940-2020. Attracted by ACE2’s accuracy, speed, and ease of use, several external groups have now published results from interesting forecast and research applications of ACE2-ERA5 beyond those that we had foreseen. This is an important step in the evolution of Ai2’s climate emulator from a research curiosity to a widely used practical tool, and helps us further improve ACE2. In this blog, we’ll dive into two of those external applications in more detail.

Seasonal forecasting

Many weather forecasting centers now use physically-based earth system models for forecasts of average weather conditions across a three-month season, which are made at least a month before the start of that season. Several groups pursuing AI-based weather forecasting are also developing such seasonal forecasts. Because oceans strongly affect weather, seasonal weather forecasts typically include forecasting how sea-surface temperatures around the world will change.

ACE2-ERA5 was designed for longer multi-decade simulations trained with specified time-varying sea-surface temperatures (SSTs). It was not designed for seasonal forecasting, and does not forecast sea-surface temperature. However it is fast to run, easy to set up, and makes accurate weather forecasts given prior knowledge of the SST. This tempted researchers at the UK Met Office, a leading international weather forecast center, to see if ACE2-ERA5 could make skillful seasonal forecasts given a simple assumption about how SST evolves during the forecast. They downloaded ACE2-ERA5 from our public HuggingFace repo and quickly got it running, with no help or knowledge from Ai2. For each of the last 20 years, they made retrospective four-month forecasts starting with their best guess at atmospheric conditions and sea-surface temperature on November 1. They used ACE2 to forecast maps of 3-month mean temperature and precipitation over the following December-January-February. They made a crude (but decent) assumption that sea-surface temperature deviations from seasonal normals remained at their observed Nov. 1 value throughout the simulation. 

To their surprise, ACE2 performed remarkably well, with comparable forecast skill to their own flagship physically-based seasonal forecast system (Figure 1) at 1000-fold less computational expense. Regions of the globe in which their forecast model was most skillful closely matched the analogous regions for ACE2. They reached out to Ai2’s Climate Modeling team to share their findings, and posted them in a paper (Kent et al. 2025) on which they invited Oli Watt-Meyer, Lead Research Scientist at Ai2 and author of the paper on ACE2, to be a coauthor. While just a feasibility test, this study suggests that future versions of ACE or similar AI models may be our best source of seasonal weather forecasts.

Indeed, we are now coupling ACE2 to a realistic AI emulator of the global ocean called Samudra, developed by colleagues at New York University and NOAA’s Geophysical Fluid Dynamics Laboratory. Our primary goal in this collaboration is to enable fast, accurate simulation of future ‘anthropocene’ climates and their potential impacts on human and natural systems, but we also hope that this will improve our emulator-based seasonal forecasts.

Emulating how ocean temperature patterns affect global warming 

Our users have also been testing the utility of ACE2 for scientifically motivated climate modeling applications, including the SST pattern effect — how natural and human-caused changes in the spatial pattern of sea-surface temperature are affecting the decade-by-decade rate of warming around the globe. This is important for using past observations of climate change to better predict climate change over the next few decades. With ongoing improvements in design and training, emulators like ACE2 promise to greatly accelerate scientific analyses that rely on doing thousands of years of global climate simulation with computationally expensive physics-based models.

In 2017, Zhou et al. proposed ‘Sea surface temperature Green’s function’ (SSTGF) analysis to interpret the SST pattern effect in climate models. Since then, this analysis has been widely adopted. It depends on how well climate models simulate cloud-forming processes (Bloch-Johnson et al. 2024), a long-standing challenge for which newly improved climate models may produce more reliable results.

The US Department of Energy recently released Version 3 of their flagship global earth system model E3SM. Over the last few months, a team of scientists performed an SSTGF analysis based on its atmospheric model component, EAMv3. SSTGF analysis requires a set of 107 pairs of global climate model simulations, each at least a decade long. In each simulation pair, the normal seasonal pattern of SST is changed in one small patch of the ocean – it is warmer in one simulation and cooler in the other. Comparing these two simulations tells us the sensitivity of the climate to changing SST just in that patch. The 107 SST patches combine to cover the world’s oceans. By this means, the climate response to any SST pattern can be broken down into more easily understood contributions from these patches. 

The 2000+ simulated years needed for SSTGF analysis of EAMv3 required many millions of CPU core-hours at a major high-performance computing center. Our collaborators at Lawrence Livermore National Labs (LLNL) suggested that we do a parallel analysis with the ACE2 emulator, which offers a huge savings of time and electrical power – only 33 hours of inference for 2200 years of inference on a single H100 GPU. No other AI climate emulator has yet done this; it requires emulating an aspect of climate called global-mean net radiation to extremely high accuracy. We trained the ACE2 emulator on customized outputs from a computationally inexpensive 50 year simulation from EAMv3 forced by historical sea-surface temperatures from 1970-2020. ACE2 is trained only to make skillful 6-hour forecasts, but by this means, long ACE2 simulations also achieve an accurate climate..

Our recently-submitted paper (Wu et al. 2025) compared the SSTGF analysis of ACE2 with the EAMv3 reference simulations. The key prediction of this analysis is a map of ‘global-mean net radiation sensitivity to SST’. Figure 2 compares maps made using the ACE2 emulator vs. those made directly from EAMv3. Overall results were encouraging - we found that ACE2 emulated the global-mean net radiative sensitivity surprisingly well, though not quite accurately enough to be used as a substitute for EAMv3. Training data spanning a more diverse range of sea-surface temperature patterns than seen in the last 50 years might greatly improve ACE2’s skill in this regard.

Another group at Colorado State University (Van Loon et al. 2025) independently published an SSTGF analysis of ACE2 trained on observations rather than a model. This complements our work – it is more philosophically satisfying but admits no rigorous way to test the resulting predictions.

Beyond these two applications, we are aware of several other external groups working with ACE2 to investigate other climate science problems such as the initiation of tropical cyclones (Chien et al. 2025) and the amplification of low-latitude global warming with height. The successes and lessons from such analyses help focus our efforts to make ACE2 better and more useful to the climate science community. Stay tuned to this blog for a future update where we’ll share how we plan to leverage ACE2 to serve another target audience: climate adaptation planners.