Probabilistic Forecasting from Weeks to Years Using AI Models

Subseasonal to seasonal (S2S) and seasonal to decadal (S2D) modeling remains a significant challenge for conventional physics-based models due to their large computational demands. Forecast systems based on artificial intelligence (AI) provide an alternative approach, with a main benefit being that AI models can create forecasts with significantly lower computational resources compared to physical models. Additional benefits include the ability to synthesize information from multiple data sources into a single model, including numerous physical models and pure observational data.

The focus of this presentation is to present an AI model that automatically calibrates posterior distributions and rapidly issues forecasts for timescales from weeks to years into the future. The model includes information about internal climate variability and climate change, which maintains its stability in an evolving climate system. The model has higher accuracy than physics-based forecasts and can be trained and issue forecasts using 1,000,000 times fewer computational resources than a physics-based model. We discuss alternative approaches for calibrating AI models for probabilistic outcomes, which is an important feature for any S2S or S2D forecast system.