Foundation Models as Assistive Tools in Hydrometeorology: Opportunities, Challenges, and Perspectives

Abstract

Most state-of-the-art AI applications in hydrometeorology are based on classic deep learning approaches. However, such approaches cannot automatically integrate multiple functions to construct a single intelligent agent, as each function is enabled by a separate model trained on independent data sets. Foundation models (FMs), which can process diverse inputs and perform different tasks, present a substantial opportunity to overcome this challenge. In this commentary, we evaluate how three state-of-the-art FMs, specifically GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro, perform across four key task types in hydrometeorology: data processing, event diagnosis, forecast and prediction, and decision-making. The models perform well in the first two task types and offer valuable information for decision-makers but still face challenges in generating reliable forecasts. Moreover, this commentary highlights the concerns regarding the use of FMs: hallucination, responsibility, over-reliance, and openness. Finally, we propose that enhancing human-AI collaboration and developing domain-specific FMs could drive the future of FM applications in hydrometeorology. We also provide specific recommendations to achieve the perspectives.