Google: How AI Meets Physics to Decode Extreme Weather

Google is tackling one of the toughest challenges in global-scale weather and climate modelling: predicting precipitation, with its unpredictable timing, location and intensity.

To address this, Google has unveiled NeuralGCM, an open-source hybrid atmospheric model that merges ML with traditional physics-based simulation.

This fusion enables faster, more precise global atmospheric forecasts at significantly reduced computational cost.

NeuralGCM delivers notable accuracy improvements across key metrics – from average precipitation and extreme rainfall to the daily weather cycle – with standout performance in modelling the most intense 0.1% of rainfall events.

Part of Google’s broader Earth AI initiative, NeuralGCM bridges the gap between physics-driven and AI-native models, extending the reach of fully AI-powered systems like the latest WeatherNext 2 to provide deeper, longer-range weather and climate insights.

Using AI for physics-related weather

According to the UK Met Office, climate change is fuelling a rise in melting sea ice and glaciers, higher sea levels, shifts in the Earth’s hydrological cycle, and more frequent heatwaves and rainfall.

To simulate precipitation accurately, Google must first understand cloud behaviour – an inherently complex challenge, given that clouds vary dramatically in size, composition and structure.

These variations make it difficult for large-scale climate models to resolve atmospheric detail at fine resolutions.

Traditionally, models approximate the influence of small-scale processes like cloud formation using parameterisations – simplified formulas built on broader variables.

NeuralGCM replaces this approach with a neural network capable of learning these effects directly from historical weather data.

In this new iteration, precipitation modelling is enhanced by training the ML component directly on satellite-based precipitation observations rather than relying on reanalysis datasets.

Earlier versions of Google’s model, like most ML-driven weather systems, were trained on reconstructed atmospheric conditions combining physics models with observational data – an approach that often missed the complexity of cloud dynamics and precipitation extremes.

To overcome this, Google’s team trained NeuralGCM’s precipitation module using NASA satellite data collected between 2001 and 2018, powered by the model’s differential dynamical core.

Unlike earlier hybrid physics and AI models limited to reanalysis or simulated outputs, NeuralGCM can now learn a more realistic, machine-learned parameterisation of precipitation informed by high-fidelity observational data – a leap forward for AI-assisted climate modelling.

“Better weather models equal better climate resilience,” says Robert Little, Sustainability Strategy Lead & Subject Matter Expert at Google.

“I’m excited to see where this takes the field of climate science next.”

How can Google forecast precipitation?

NeuralGCM’s capabilities were benchmarked using WeatherBench 2, evaluating two-week forecasts against a leading physics-based model from the European Centre for Medium-Range Weather Forecasts (ECMWF).

Tested on forecasts initialised at noon and midnight throughout 2020 – with this data excluded from training – NeuralGCM consistently outperformed the ECMWF model at low resolution across most precipitation metrics.

It demonstrated superior accuracy in both 24-hour and 6-hour accumulated rainfall across all 15 forecast days, with particularly strong performance over land, where the impacts on people and ecosystems are most significant.

While its current 280km resolution remains too coarse for operational weather prediction, the findings highlight clear potential for scaling this AI-physics hybrid approach to higher-resolution forecasting, signalling an important step forward for data-driven climate simulation.

Over longer timescales spanning years to decades, NeuralGCM also demonstrated strong skills in: 

• Reproducing average and extreme precipitation patterns
• Achieving an average mean error of less than half a millimetre per day
• Reducing error by 40%  compared with leading global atmospheric models used in the latest Intergovernmental Panel on Climate Change report, with even larger improvements over land. 

The model delivered significant improvements in detecting extreme rainfall, especially within the most intense 0.1% of precipitation events.

This advancement tackles persistent challenges in traditional physics-based models, which often overpredict light rain while underrepresenting heavy downpours – a critical gap NeuralGCM’s AI-driven approach helps bridge.

NeuralGCM also captured the daily timing and intensity of rainfall with greater accuracy, successfully modelling strong diurnal cycles such as afternoon precipitation over the Amazon during summer.

This represents a marked improvement over conventional climate models, which often simulate rainfall too early in the day.

Precisely modelling when and where precipitation occurs is vital for real-world applications – from flood and drought management to climate science, ecosystem resilience and public safety – underscoring the broader societal value of advancing AI-driven weather prediction.

The future of AI

“We believe this is a step forward for large-scale precipitation forecasts and simulations, and we already have early support in the real world,” Google says.

The Intergovernmental Panel on Climate Change (IPCC), in its Sixth Assessment Report (2021), found that human-driven increases in greenhouse gases have intensified the frequency and severity of extreme weather events – including monsoons, droughts and heatwaves.

In response to these growing challenges, the University of Chicago partnered with the Indian Ministry of Agriculture and Farmers Welfare to leverage Google’s NeuralGCM in predicting the onset of the monsoon season.

During summer 2025, the collaboration selected NeuralGCM, alongside another advanced model, to develop and deploy an AI-powered forecasting tool, aiming to deliver more accurate and timely monsoon predictions for climate resilience and agricultural planning.

“Since introducing NeuralGCM we have made everything available as open-source code on which we hope people can build,” the company says.

“This precipitation model is also being openly released to the extended community.

“Ultimately our hope is that these efforts will bring us one step closer to accurate long-term projections of future precipitation, especially under climate change.”