AI-powered weather and climate models are set to transform the future of forecasting, providing more accurate and cost-effective predictions. 

By integrating traditional weather modelling with machine learning, a groundbreaking model promises to deliver high-quality forecasts at a fraction of the cost.

In a study published in Nature, researchers from Google, MIT, Harvard, and the European Centre for Medium-Range Weather Forecasts introduced NeuralGCM, a hybrid model combining physical simulations with machine learning. 

This approach offers significant “computational savings” and enhances simulations crucial for understanding and predicting the Earth's systems.

NeuralGCM represents a new era in weather and climate prediction. 

As a “general circulation model”, it mathematically describes the physical state of Earth's atmosphere and solves complex equations to forecast future conditions. 

Unlike traditional models, NeuralGCM incorporates machine learning to better understand and predict less well-known physical processes, such as cloud formation, ensuring consistency with the laws of physics.

This hybrid model can generate weather forecasts days and weeks in advance and predict climate trends months and years ahead. 

In standardised tests using WeatherBench 2, NeuralGCM's performance was comparable to other advanced models for short-term forecasts and matched the best traditional models for longer-term predictions. 

Additionally, it successfully forecasted rare weather phenomena, including tropical cyclones and atmospheric rivers.

Machine learning models rely on algorithms that identify patterns in vast datasets to make predictions. 

Given the complexity of climate and weather systems, these models require extensive historical and satellite data for training, which is computationally expensive. However, once trained, these models can generate predictions quickly and affordably.

This high training cost but low usage cost model is similar to other machine learning systems, such as GPT-4, which took months to train at over $100 million but can respond to queries almost instantly.

One significant advantage of NeuralGCM is its improved ability to generalise beyond its training data, thanks to its physics-based core. 

This feature is critical as climate change brings more unprecedented weather conditions. However, machine learning models can struggle in novel situations, and ongoing research is essential to ensure they remain reliable as the climate evolves.

Currently, no machine learning-based weather models are used for daily forecasting. However, the field is rapidly advancing, suggesting that future forecasts will increasingly rely on machine learning.