Accurate precipitation forecasting is vital for flood and drought management, climate science, ecosystem resilience, and public safety, with ambitions to scale this approach to operational use in the future.

NeuralGCM is a hybrid atmospheric model that blends machine learning with physics-based simulation to enhance global weather forecasts.

Early results show mean daily error shrinking to under 0.5 mm and a roughly 40% reduction in error versus leading global models used in IPCC assessments.

Over multi-year timescales, NeuralGCM reliably reproduces precipitation patterns and lowers errors, addressing biases that overestimate light rain and underestimate heavy rain.

Evaluations with WeatherBench 2 against ECMWF at 280 km show NeuralGCM outperforming the physics-based model on many precipitation metrics, with stronger gains over land but current resolution limits operational use.

Benchmark results also show NeuralGCM beating a leading ECMWF model at low resolution on several precipitation metrics, including 24-hour and 6-hour rainfall, especially over land.

While current resolution limits operational forecasting, the approach holds promise for higher-resolution forecasts and long-term climate simulations spanning years to decades.

NeuralGCM improves long-range projections by better capturing average and extreme precipitation and crucially enhances the timing of intense rainfall events in the top 0.1%.

The model achieves under-0.5 mm mean daily error and improved detection of the most intense 0.1% rainfall events, even at the 280 km resolution.

NeuralGCM more accurately models daily rainfall timing and diurnal cycles, such as Amazon afternoon showers, mitigating common biases in traditional models.

The precipitation module is trained on NASA satellite data from 2001 to 2018 using a differential dynamical core to learn cloud-process parameterisations, reducing reliance on conventional methods.

Google is releasing NeuralGCM as open-source to foster community-driven development and improve long-term precipitation projections amid climate change.