Ai2 Collaborates with Google to Develop Climate Models: How Is AI Studying the World Decades from Now?

Ai2 has received a three-year grant from Google.org and will develop NeuralGCM+ with technical support from Google Research to explore faster long-term climate simulations. To understand this collaboration, it is important to distinguish between weather forecasting and climate projections: the former tracks short-term changes, while the latter examines long-term risks under various scenarios. Speed, physical constraints, ocean coupling, and uncertainty collectively determine whether a model can truly support decision-making.

On November 4, 2026, the Allen Institute for Artificial Intelligence (Ai2) announced the launch of a new climate modeling project: Google.org is providing a three-year grant in the millions of dollars, and Google Research is offering technical support to advance the development of NeuralGCM+. The project focuses on climate trends spanning several years to several decades and is still in the research and development phase.Ai2 Partnership Announcement

What makes this collaboration noteworthy is that it broadens the scope of AI research: beyond simply predicting whether it will rain in a few days, people also need to know what challenges water resources, heat waves, and infrastructure will face in the coming decades. Answering the latter type of question requires a different set of evaluation criteria.

Weather Forecasts and Climate Projections Answer Different Questions

Weather forecasts use recent observations to predict how the atmosphere will change in the near future. Climate projections, on the other hand, examine how the climate system responds to long-term external changes under a set of reasonable assumptions. One focuses on whether a specific rainfall event will occur, while the other concerns the long-term distribution of rainfall and its changes. Climate projections on a scale of several decades should not be interpreted as a daily weather calendar for a specific year or day in the future.Australian Climate Change Platform: Projections and Confidence Levels

Consider this real-life example: When planning a weekend trip, you need a short-term weather forecast; when designing a reservoir that will be in use for decades, you need to consider long-term risks associated with water inflow. Even if the latter scenario cannot provide the exact dates of every rainfall event, it can still help designers compare the resilience of different design options. As the time scale changes, so does the information that is truly useful.

Therefore, the fact that a model “can simulate several decades” does not in itself indicate that it is more accurate. We must also ask: Is it simulating global average temperature, regional precipitation, or extreme events? These variables present different levels of difficulty, and their validation methods cannot be treated interchangeably.

What Does AI Do? What Do Physical Equations Do?

NeuralGCM takes a hybrid approach: it uses physical equations to model large-scale atmospheric motion and neural networks to represent smaller-scale processes. Traditional models divide the Earth into grids, but processes such as clouds and precipitation often occur on a smaller scale than the grid cells and cannot be solved directly on a cell-by-cell basis, requiring approximate representations. Here, AI takes on the task of learning these processes.Google Research Technology Overview

This is fundamentally different from having a chatbot “guess the weather” based on textual descriptions. Simulation requires step-by-step forward calculation, and local processes can in turn affect the overall state. Physical structures provide constraints, while learning systems supplement the parts that are difficult to calculate in detail; the collaboration between the two is the focus of this research.

However, incorporating physical equations does not guarantee accuracy. Google’s research also points out that feedback between small-scale and large-scale processes accumulates over time; if these connections are ignored during training, numerical stability issues may arise. Even if the model performs well in the short term, it still needs to be tested to ensure it operates reasonably over the long term.Google Research: Notes on Joint Training and Stability

From the user’s perspective, a “hybrid model” should be a design choice that requires validation, rather than a label that automatically confers credibility. A smooth-looking global animation merely demonstrates that the model is running; it cannot replace tests for error and stability.

The shift from the atmosphere to the ocean is key to this collaboration.

The Ai2 project plans to couple NeuralGCM with ocean and sea ice models to conduct stable, decades-long simulations and evaluate the model’s performance under warmer climate conditions not encountered during training. The team also plans to make the model, modeling framework, and documented workflows publicly available, and to improve performance on commonly used GPUs. These are research and development goals and should not be presented as capabilities that have already been achieved.Ai2 Announced Research Areas

The ocean absorbs and redistributes heat, influencing long-term climate. One assessment of this collaboration is that the challenge lies not only in adding a module to existing models, but also in whether the coupled system can operate in a coordinated manner over the long term. Progress made in individual atmospheric models does not mean that the problem of atmosphere-ocean coupling has been solved.

Similarly, “planned disclosure” must be examined on a case-by-case basis. Researchers are concerned not only with the availability of code, but also with whether they can reproduce the experiments, retrain the model, understand the inputs and outputs, and determine whether the terms of use are suitable for their own work. The specific version and license of future results will be subject to the actual content of the release.

Running Faster Can Make Uncertainty Clearer

There are various sources of uncertainty in climate projections: humanity has yet to determine how future emissions will change, and models’ representations of the Earth system are inevitably subject to bias. These two types of limitations must be examined through different scenarios and multiple models.Canadian Climate Service Center: Scenarios and Models In addition, the climate itself is subject to natural variability; the Australian Climate Change Platform also identifies natural variability as a source of uncertainty that must be taken into account when using climate projections.Australian Climate Change Platform: Uncertainty and Confidence

Acceleration cannot determine humanity’s future emissions trajectory, but it holds the promise of reducing the cost of exploring more scenarios, different starting points, and different models. This is a particularly promising application of AI simulation: conducting more thorough comparisons that were previously too costly or insufficient to perform, rather than presenting a single curve as the only answer.

The Canadian Climate Service Center also emphasizes that even when using the same emissions scenario, different models will still produce different results; therefore, it combines multiple models into a ensemble to illustrate the range of outcomes. When looking at such charts, readers should consider both the central trend and the range of variation, rather than focusing solely on the most prominent line.Canadian Climate Service Center: Scenarios and Models

There is still a gap between research findings and local decision-making.

There are scale differences between the grid of global models and specific cities or watersheds. Topography can influence local temperatures, and extreme rainfall events may occur over smaller areas. Studies of actual impacts often require downscaling to convert coarser-resolution global data into finer-resolution regional data.CanadianStatement from the Climate Service Center Regarding Downscaling

Therefore, advances in global modeling should not be directly translated into precise projections of water shortages for a specific city. For local users, a more practical question might be: If conditions turn out to be drier than expected, will the original plan still work? Which investments can be implemented in phases? What indicators should prompt adjustments to the plan?

When reporting research findings, it is also important to distinguish between “simulating the past” and “projecting the future.” The ability to accurately reproduce historical data is a crucial test, but when faced with a warmer environment that differs from the training data, new tests are needed. Just because a model passes a particular historical test does not mean that all regional results for the coming decades can be considered equally reliable.

One detail worth examining is the comparison criteria: What input data did the model use? On what type of hardware was it run? Were the spatial resolutions comparable? Was the comparison based on average error or extreme cases? Only when these conditions are clearly specified do speed and accuracy become meaningful. For the general reader, these questions are more helpful than simply memorizing a striking speed-up factor.

In further evaluating NeuralGCM+, three key points should be considered: whether the long-term simulations are stable, whether the model is reliable under unprecedented climate conditions, and whether external teams can reproduce and verify the results. Its value ultimately lies in its verifiable research capabilities: enabling people to more thoroughly compare possible futures and, based on that, make more informed choices.