Hot AI News

AI climate emulators: How a learned model replaces a slow climate simulation

An AI climate emulator learns from past weather and simulations to mimic the atmosphere, running years of climate in minutes. How it works, who builds it and where it breaks.

AI climate emulators: How a learned model replaces a slow climate simulation
Photo illustration: Aiholics

An AI climate emulator is a neural network trained to imitate how the atmosphere evolves, so it can produce years of simulated weather in minutes instead of weeks. It does not solve physics equations. It learns the patterns from data that a physics-based model or a reanalysis (a record of past weather) already produced.

Think of a pilot who has logged thousands of hours in a very slow, very accurate flight simulator. Eventually the pilot can predict what the simulator will do next without running its full calculations. An emulator does the same for the atmosphere.

Learning the answers of a physics model instead of its equations

A traditional climate model splits the atmosphere into a grid of boxes and solves equations for heat, moisture and wind in each box, step by step. It needs supercomputers, and long runs can take weeks.

An emulator skips the equations. It is given the state of the atmosphere at one moment (temperature, humidity, winds, pressure) and learns to predict the state a few hours later. The training examples come from a physics model or from ERA5, the fifth-generation reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF). According to the Copernicus Climate Data Store, ERA5 covers 1940 to the present, hourly, at 0.25-degree resolution for the atmosphere, under a CC-BY license.

Advertisements

Stepping forward in small time steps, thousands of times

A sloping research display shows five successive maps of a cloud system evolving over the North Atlantic.
Photo illustration: Aiholics

The emulator is autoregressive: its output becomes its next input. To simulate a year, it takes a step, feeds the result back in, and repeats. Ai2’s ACE2 uses 6-hour steps, so one year is about 1,460 steps.

Small errors can pile up over so many steps, and the model can drift into an unphysical state. This is the main engineering problem, and why climate emulators differ from weather models, which only need to be good for about two weeks. Ai2’s ACE2 paper says its model conserves global dry air mass and moisture exactly, which helps it stay stable. It does not conserve energy exactly, the authors note.

Outside forces also matter. ACE2 takes sea surface temperature and carbon dioxide as inputs, so the model can respond to a warming ocean instead of only repeating the past.

From one fixed answer to a spread of plausible ones

The atmosphere is chaotic, so scientists run many slightly different simulations, called an ensemble, and look at the spread. A model trained to minimize average error produces blurry output, because the average of many plausible states is smooth. Fine detail such as storm rainfall gets smeared out.

Newer emulators are stochastic: random noise is part of the model, so each run gives a different but plausible result. They are trained with a score that rewards both accuracy and honest spread. One such score is the continuous ranked probability score (CRPS). A calibrated ensemble is one whose spread matches its real error, so a wide spread signals real uncertainty.

Google DeepMind’s GenCast, described in 2023, took this route for weather. Its authors report an ensemble of 15-day forecasts at 0.25 degrees in 8 minutes, with more skill than ECMWF’s operational ensemble on 97.4% of 1,320 targets.

Ai2’s ACE2.5-ERA5 applies the idea to climate. According to its model card, the main model has 46M parameters, down from 456M in ACE2-ERA5. The card says ensemble spread matches error to within 4% from day 3. It also says the model keeps 86% of the fine-scale spectral power in ERA5 precipitation, against 36% for its predecessor. These figures are Ai2’s own and were not independently verified.

Advertisements

Where emulators came from

Ai2 published the first ACE paper in October 2023 (Watt-Meyer and colleagues, with co-authors including NVIDIA researchers). It was a 200M-parameter model that imitated a 100-km atmospheric model called FV3GFS. The authors reported about one second of wall-clock time per simulated day on one NVIDIA A100 GPU, against roughly 77 seconds for the reference model on 96 processor cores. It stayed stable for 100 years, though it was trained on a fixed seasonal cycle of ocean temperatures.

ACE2 followed in November 2024. It has 450M parameters, 1-degree resolution and eight vertical layers, and was trained on ERA5 and on a simulation from the NOAA GFDL SHiELD model. The authors report about 1,500 simulated years per day. They say it reproduces tropical cyclones, the Madden-Julian Oscillation (a slow tropical rain pattern) and sudden stratospheric warmings, and tracks 80 years of observed trends.

Others work on the same problem. NeuralGCM from Google mixes a neural network with a conventional physics core, and independent studies on arXiv compare it with ACE2.

What an emulator cannot tell you

An emulator is only as good as its training data. If ERA5 or the source simulation has biases, the emulator inherits them. It also tends to be reliable only within conditions it has seen. The ACE2 authors write that its responses to changing sea surface temperature and CO2 on their own are “not entirely realistic”, and that they do not expect it to simulate strong climate change properly.

Ai2 labels the ACE models research tools that should not be used for operational climate predictions. Other limits come from the design: 1-degree grids cannot resolve single thunderstorms, and the ACE2 paper says its medium-range weather skill trails dedicated forecast models.

The practical gain is speed. When a run costs minutes, researchers can test thousands of scenarios, run large ensembles and study rare events that a supercomputer budget would not allow.

Watch for three things: emulators trained on more varied forcing data, coupling to the ocean and sea ice (the original ACE authors named this as a needed step), and independent tests of whether ensemble spread and rainfall statistics hold up outside the training years.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Daniel ReedAI Research, Safety & Ethics Analyst

Daniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like