Google Research unveils NeuralGCM model

Google Research has published a new explainer on NeuralGCM, a hybrid climate model that combines a traditional atmospheric simulator with machine learning to model global precipitation more accurately. The post, published on January 12, 2026, said the system was trained on NASA precipitation observations and was designed to improve long-range rainfall simulation, including the daily precipitation cycle and extreme events. (research.google) Google tied the release to a related research publication on precipitation modeling and to publicly available code and model resources. ### What exactly did Google release? Google Research said on January 12 that it had published details of NeuralGCM’s precipitation modeling work in a blog post by research scientist Janni Yuval. (github.com) The post described NeuralGCM as a hybrid system that pairs physics-based modeling with a neural network trained on observational precipitation data. Google Research’s publication page for the underlying work described the model as a “neural general circulation model” for precipitation. That page said the system was built on the differentiable NeuralGCM framework and was stable for decadal simulations. (research.google) ### What makes NeuralGCM different from a standard climate model? NeuralGCM combines a fluid-dynamics solver with learned components for smaller-scale atmospheric processes, according to Google Research’s earlier documentation of the framework. Google said that setup lets the model retain physical structure while using machine learning to represent processes that are harder to capture with conventional parameterizations alone. (research.google) The January 12 Google Research post said the precipitation version was trained directly on NASA precipitation observations. Google said that observational training helped the model simulate rainfall more accurately than other approaches, particularly for the diurnal cycle and extremes. (research.google) ### What did Google say the model improved? Google Research said the model showed reduced biases, a more realistic precipitation distribution, improved representation of extremes and a more accurate diurnal cycle. The Google Research publication page also said the system outperformed existing general circulation models, ERA5 reanalysis and a global cloud-resolving model in simulating precipitation. (research.google) The same publication page said the model outperformed the ECMWF ensemble for mid-range weather forecasting. Google’s broader NeuralGCM research page separately said the framework was competitive with machine-learning models for one- to ten-day forecasts and with the European Centre for Medium-Range Weather Forecasts ensemble for one- to fifteen-day forecasts. (research.google) ### How detailed is the model? Google Research’s precipitation paper page said the model runs at 2.8-degree resolution. Google’s earlier NeuralGCM work said other versions of the framework had been trained at 0.7-degree, 1.4-degree and 2.8-degree resolution using ECMWF weather data from 1979 to 2019. (research.google) A public Colab notebook linked from the project documentation said users can load a pre-trained intermediate 1.4-degree deterministic model by default. The notebook also listed deterministic 0.7-degree and 2.8-degree checkpoints, stochastic 1.4-degree variants and 2.8-degree stochastic models that predict precipitation. (research.google) ### Where can researchers find the code and models? GitHub shows NeuralGCM as a public Python library for building hybrid machine-learning and physics atmospheric models for weather and climate simulation. The repository says the code is licensed under Apache 2.0 and trained model weights are available under a Creative Commons Attribution-ShareAlike 4.0 license. (research.google) The project documentation says the codebase includes components for reproducing and extending results from the NeuralGCM paper, along with installation guides, API materials and inference notebooks. A documentation page also directs users to GitHub for support and lists a team contact email. ### What comes next for NeuralGCM? (colab.research.google.com) Google’s public repository shows the NeuralGCM project remains active, with updates logged in 2025 and 2026 and documentation pointing users to model checkpoints and example workflows. The latest public release listed on GitHub was version 1.2.2 on July 31, 2025, while the repository and docs continue to host notebooks and datasets for users testing the models. (github.com) (neuralgcm.readthedocs.io) (github.com)