Google researchers are testing a middle path in weather prediction: not replacing traditional atmospheric science with artificial intelligence, but combining the two. Their new model, NeuralGCM, is designed to use machine learning where it helps most while keeping the physics-based structure that has long guided forecasting.
The result could matter beyond daily forecasts. According to the researchers and outside experts cited in the source article, the larger opportunity may be in climate and extreme weather modeling, where conventional systems can be costly and slow to run at scale.
Why weather forecasting is split between two approaches
Weather prediction has been shaped for decades by general circulation models. These systems use complex equations to represent changes in the atmosphere. They have dominated the field for the last 50 years because they can produce accurate projections, but they are also slow and expensive to run.
Machine-learning weather models take a different route. They learn patterns from years of past data and can produce forecasts very quickly. That speed and efficiency have made them attractive, especially as AI systems improve.
But the tradeoff is important. The source article says machine-learning techniques can struggle with long-term predictions, while conventional general circulation models remain strong but computationally heavy. That has left experts divided over which approach will be most reliable in the years ahead.
NeuralGCM is Google’s attempt to avoid choosing one side. As Stephan Hoyer, an AI researcher at Google Research and a coauthor of the paper, puts it: “It’s not sort of physics versus AI. It’s really physics and AI together.”
How NeuralGCM combines machine learning with physics
NeuralGCM still uses a conventional model to calculate some of the large atmospheric changes needed for a forecast. That matters because large-scale physics remains central to modeling how the atmosphere evolves.
The AI component is added more selectively. According to the source article, it is used in areas where larger models often have weaker performance, especially at scales smaller than about 25 kilometers. Those smaller-scale features can include cloud formations or regional microclimates, such as San Francisco’s fog.
Hoyer describes the role of AI this way: “That’s where we inject AI very selectively to correct the errors that accumulate on small scales.”
This structure is the key idea behind NeuralGCM. It does not ask machine learning to handle the entire forecasting problem on its own. Instead, it lets conventional modeling handle broad atmospheric behavior while AI helps correct smaller-scale errors that can build up inside the system.
What Google says the model can do
The researchers say NeuralGCM can produce quality predictions faster and with less computational power. They also say it is as accurate as one-to-15-day forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), which is a partner organization in the research.
That comparison matters because ECMWF is presented in the source article as a benchmark for medium-range forecasting. If NeuralGCM can match that level while using less computational power, it could offer a practical way to make advanced modeling more accessible.
The source article also highlights how compact AI-based systems can be once they are trained. Machine-learning models are typically trained on 40 years of historical weather data from ECMWF. Google’s GraphCast, for example, can run on less than 5,500 lines of code, compared with nearly 377,000 lines required for the model from the National Oceanic and Atmospheric Administration, according to the paper.
Those figures do not mean code length alone determines forecast quality. But they do show why researchers are interested in AI-based weather prediction: smaller, faster systems may be easier to run repeatedly and at lower cost.
The bigger prize may be climate risk, not local weather
Aaron Hill, an assistant professor at the School of Meteorology at the University of Oklahoma, was not involved in the research. He argues that the most important promise of NeuralGCM-like systems is not better local forecasts.
Instead, he points to larger-scale climate events that are expensive to model with conventional approaches. The possibilities described in the source article include predicting tropical cyclones with more notice and modeling more complex climate changes that are years away.
Hill explains the bottleneck clearly: “It’s so computationally intensive to simulate the globe over and over again or for long periods of time.” That cost limits what researchers can do with even the best climate models.
Hoyer also says short-term weather prediction has been useful for validating NeuralGCM, but the broader goal is longer-term modeling, especially for extreme weather risk. In that sense, daily forecasting is partly a proving ground for a system that may later be more useful in climate research.
Who could use NeuralGCM next
NeuralGCM will be open source, according to the source article. Hoyer says he looks forward to climate scientists using it in their research, but the potential audience is wider than academia.
The article identifies several groups that may care about high-resolution predictions:
- Commodities traders, who pay for detailed weather information.
- Agricultural planners, who also value high-resolution forecasts.
- Insurance companies building products such as flood or extreme weather insurance.
The insurance use case is especially tied to the problem of climate change. The source article says models used by insurance companies are struggling to account for its impact.
Still, the field is moving quickly. Hill says many AI skeptics in weather forecasting have been won over by recent developments, but the pace of new releases is hard for researchers to absorb. He notes that it can seem as if a new model is released by Google, Nvidia, or Huawei every two months.
That speed creates a practical problem. Researchers need time to understand which tools are useful and to apply for research grants accordingly. Hill sums up the mood this way: “The appetite is there [for AI].” But he adds, “But I think a lot of us still are waiting to see what happens.”
For now, NeuralGCM’s significance is not that it settles the debate between AI and traditional forecasting. Its importance is that it reframes the debate. The future of weather prediction may depend less on choosing between machine learning and physics, and more on deciding where each one belongs inside the same model.