NIVA: Multimodal Foundation Model for Earth System Intelligence

Current AI models for weather and climate forecasting have improved skill and reduced computational costs, but they often struggle to model the complex, coupled dynamics of the entire Earth system. This limitation restricts their ability to extend predictions beyond a typical two-week horizon. To address this, researchers introduce NIVA, a multimodal foundation model. NIVA is designed to learn unified representations across various Earth system components, such as the atmosphere, ocean, ice, and land. As a proof of concept, the initial validation focuses on a two-modality setting (ocean and atmosphere). By training on extensive Earth system simulations, NIVA successfully captures key modes of climate variability and demonstrates accurate prediction of major climate indices. This foundational work paves the way for improved subseasonal-to-seasonal prediction and more comprehensive Earth system intelligence.