Decision-oriented benchmarking of AI weather models
- AI weather models now often beat physics-based models on standard metrics at a fraction of the computing cost, but they are judged on aggregate scores rather than on the decisions people face.
- We introduce a decision-oriented benchmarking framework linking meteorology, AI and the social sciences, and apply it to a 150-year-old problem: forecasting Indian monsoon onset for rain-fed agriculture.
- AI models predict onset skillfully weeks ahead. The framework informed a 2025 government effort that sent AI-based forecasts to 38 million farmers, which captured an unusual weeks-long pause in the monsoon.
Reference
Masiwal, R. et al. Decision-oriented benchmarking to transform AI weather forecast access: Application to the Indian monsoon. arXiv (2026). arXiv:2602.03767
Aitken, C., Masiwal, R. et al. Designing probabilistic AI monsoon forecasts to inform agricultural decision-making. arXiv (2026). arXiv:2603.07893


