Physically-based climate models discretize equations representing individual time-evolving processes like clouds, rain, wind, land and sea-ice, and ocean currents on a computational grid. Some processes (e.g. clouds) are less reliably encoded than others (e.g. winds). Hence such climate models have biases in representing present-day climate and produce an undesirably large range of projections of future climate change for a given human-caused change in CO2 or other climate forcings. Current AI-based climate models reduce present-day climate bias, but don’t generalize well to future climate change.
We are starting a 3-year project funded by google.org to develop an open source ‘AI-first’ climate model that can demonstrably project future climates more accurately than current physically-based climate models when trained on historical observations by judicious design of the AI to incorporate physical principles that reliably apply in any climate, seen or unseen. The model should be lightweight and easy for an ML-savvy climate scientist to train and deploy.
Our starting point is the Google Research NeuralGCM model, a hybrid ML model that uses a conventional discretization of winds encoded in JAX and learns a column-local representation of all other atmospheric and land processes from historical data. Compared to other AI climate models, NeuralGCM uses ML in a conceptually simpler way and shows somewhat better future-climate generalization skill. It is an open-weight model but its training pipeline and documentation need to be adapted for a broader user community.
You will join a small team to develop the desired NeuralGCM+ model and foster its uptake by the climate modeling community. This includes coupling it to an ML-based ocean model and designing approaches to demonstrably improve its skill in unseen climates. This position has a three year term; there is potential for longer-term funding if the NeuralGCM+ project demonstrates exceptional progress.