This presentation explores the opportunities and limitations of artificial intelligence and machine learning (AI/ML) in climate modelling, with particular emphasis on the distinction between weather prediction and climate projection. It introduces a framework for discussing AI applications: AI-Before (pre-processing inputs), AI-Inside (replacing parts of traditional models), AI-Outside (post-processing and analysis), and AI-Instead (fully data-driven models). While significant opportunities exist in the first three categories, the presentation argues that fully replacing physical climate models with AI remains problematic.

A central theme is that climate science must address futures that have not yet been observed. Unlike weather forecasting, climate projections must account for tipping points, novel states, and complex nonlinear behaviour that are absent from historical observations, making purely data-driven approaches difficult to train and validate.

The presentation reviews emerging AI-based climate models such as NeuralGCM and ACE2, acknowledging their impressive capabilities while noting that their strongest results often rely on training data generated by traditional climate models. It argues that future progress will depend on combining AI with physics-based simulation, using AI for emulation, downscaling, ensemble generation, and process representation. Ultimately, robust climate science will continue to require physical models capable of exploring unobserved futures, with AI acting as a powerful complement rather than a complete replacement.

Presentation: pdf (1.5 MB)