Google has introduced NeuralGCM, a new open-source atmospheric model that blends physics-based climate modelling with artificial intelligence to improve global precipitation forecasting, including extreme rainfall events that have long challenged traditional systems. The development sits within Google’s broader Earth AI initiative and reflects growing interest in using hybrid AI approaches to strengthen climate resilience.

Precipitation remains one of the hardest variables for global climate and weather models to simulate accurately. Rainfall is highly localized, episodic, and influenced by small-scale processes such as cloud formation, which conventional physics-based models struggle to resolve. NeuralGCM addresses this challenge by combining large-scale fluid dynamics with machine learning systems trained directly on real-world observations.

 

From Parameterisations to Learning From Reality

 

Conventional climate models rely on “parameterisations” to approximate the effects of small-scale atmospheric processes like clouds and convection. These approximations often lead to systematic errors, including overestimating light rainfall and underestimating heavy downpours. NeuralGCM replaces much of this guesswork with neural networks that learn these effects directly from data.

Unlike earlier AI-driven weather systems trained primarily on reanalyses, NeuralGCM’s precipitation component was trained using satellite-based rainfall observations collected by NASA between 2001 and 2018. This allows the model to capture cloud physics and precipitation dynamics more accurately, particularly for extreme events and daily rainfall cycles.

According to Google, this approach enables the model to better reproduce average precipitation patterns, intense rainfall, and diurnal cycles, such as afternoon rain peaks in tropical regions.

 

Read more: SkyFi Raises $12.7 Million to Turn Satellite Imagery Into Actionable Insights

 

Benchmarking Against Leading Climate Models

 

NeuralGCM’s performance was evaluated using WeatherBench 2 and compared with a leading physics-based model from the European Centre for Medium-range Weather Forecasts. Across two-week forecasts, NeuralGCM outperformed the ECMWF model at low resolution on most precipitation metrics, including 6-hour and 24-hour accumulated rainfall over a full 15-day forecast period.

The strongest gains were observed over land, where accurate rainfall forecasting has the greatest implications for people, ecosystems, and infrastructure. Over longer timescales spanning years to decades, the model reduced average precipitation error by around 40 percent compared with leading global atmospheric models referenced in recent Intergovernmental Panel on Climate Change assessments.

NeuralGCM also demonstrated notable improvements in capturing the most extreme rainfall events, particularly the top 0.1 percent of precipitation, an area where traditional models have consistently fallen short.

 

Why This Matters for Climate Resilience

 

Accurately predicting when, where, and how much rain will fall is central to managing climate risks, from floods and droughts to food security and disaster preparedness. Google says NeuralGCM has already seen early real-world validation, including use in a partnership between the University of Chicago and India’s Ministry of Agriculture to predict monsoon onset.

Robert Little, Sustainability Strategy Lead at Google, has described the work as a step toward stronger climate resilience, arguing that better weather and climate models are essential as extreme events intensify under climate change.

 

Explore OneStop ESG Marketplace: AI (Artificial Intelligence)

 

Open Source and the Road Ahead

 

Google has released NeuralGCM and its precipitation module as open-source tools, inviting researchers and institutions to build on the model. While its current resolution is too coarse for operational forecasting, the results suggest strong potential if applied at finer spatial scales.

Looking ahead, Google says its goal is to support more accurate long-term projections of future precipitation under climate change, complementing AI-only systems like WeatherNext while extending modelling capabilities across seasonal to multi-decadal horizons.