Des experts: Comment les scientifiques du climat utilisent-ils l'intelligence artificielle ?

AI weather prediction models are demonstrating capabilities that challenge long-held assumptions regarding the opacity of artificial intelligence systems. According to experts consulted by Carbon Brief, these models can deliver accurate forecasts while consuming significantly less computational power than traditional numerical simulations. The technology learns from massive datasets to predict conditions hours ahead, then iteratively feeds those outputs back into itself to extend projections up to two weeks into the future. Dr. Doug McNeall, science lead in data-driven climate modelling at the UK Met Office, noted that this efficiency marks a major shift in meteorological tools. However, he cautioned that current AI models excel primarily on short timescales where abundant observational data allows for rigorous performance evaluation. Climate projection differs fundamentally because it requires estimating future states outside the historical record. McNeall suggested a blended approach, using AI to emulate or accelerate existing physics-based models, could make climate risk information more accessible and affordable while retaining the scientific principles that underpin public confidence. Prof. William Collins, a professor of AI for climate science at the University of East Anglia, described a renaissance in the field. He highlighted that the European Centre for Medium-Range Weather Forecasts is now deploying an AI system alongside its traditional methods for official communications to the European community. Collins stated that these systems have reached a level of competitiveness with leading numerical weather prediction tools. Yet, he identified a central challenge: machine learning relies on past data, whereas climate science attempts to predict a future for which no observations exist. Dr. Amirpasha Mozaffari, a researcher in Earth artificial intelligence at the Barcelona Supercomputing Centre, pointed out specific limitations. While AI performs well on typical weather patterns, it underperforms compared to physics-based models during record-breaking or unprecedented extreme events. These extremes are critical for adaptation planning. Mozaffari indicated that solving this gap might require hybrid modeling strategies. Beyond forecasting, climate scientists apply AI to analyze satellite imagery and monitor energy grids. Despite these advances, the sector faces distinct risks. Experts cited concerns over overreliance on unverified results, the intensive energy consumption required to run these models, and the potential for AI to exacerbate existing inequalities within the climate science community. The integration of these technologies continues to evolve rapidly, prompting ongoing debate among researchers about the optimal balance between data-driven speed and physical accuracy.