Recipient of the Google.org Impact Challenge: AI for Science, Inria Chile receives US$2.5 million to accelerate an AI-powered digital twin of the ocean

Inria, the French National Institute for Research in Digital Science and Technology, announced today that its center in Chile, Inria Chile, is a recipient of the Google.org Impact Challenge: AI for Science. Inria Chile will thus receive US$2.5 million to build a multimodal foundation world model of the ocean, the basis of a digital twin that will allow us to better understand Earth’s last great unknown.

Inria Chile is one of 17 organizations selected globally from more than 3,000 applicants across 30 countries to share an updated total of US$36 million in funding. It is the only institution from Latin America among the recipients, alongside leading universities and research institutions from around the world (Find out which projects won at https://goo.gle/gic-ai-science).

The ocean regulates the climate and feeds billions of people, yet it remains the least observed system on Earth, and its data – satellites, floats, biological and genomic surveys – remain scattered in mutually incompatible formats. Scientists still cannot reliably connect the warming and acidification of the ocean to their effects on marine life: harmful algal blooms, coral bleaching, the collapse of fisheries. MOFOMO (Multimodal Ocean Foundation Models for an Ocean Digital Twin) will build the first open multimodal foundation world model that learns at once from the ocean’s physical, biological and genomic observations, coupled with physics-based models for reliable and explainable forecasts.

In its first six months, supported by the Google.org Accelerator, MOFOMO will deliver several foundational elements:

• the programme’s governance framework and the data-sharing agreements with its partners;
• version 0.1 of the open multimodal ocean data lake;
• the specification of the public benchmark, together with baseline models;
• the advisory group for the two pilots, with its first consultations completed;
• the architecture of the first version of the multimodal foundation world model, with its training pipeline running at scale.

By the end of the 36 months, it will have delivered:

• the full, harmonised data lake;
• the public benchmark against which any ocean AI model can be compared;
• the MOFOMO multimodal foundation world model itself, pre-trained at scale and released with its weights, code and documentation under permissive licences;
• a hybrid version coupled with physics-based ecosystem models, with calibrated uncertainty;
• two early-warning services validated with their end users, for harmful algal blooms with Chile’s aquaculture sector and for coral bleaching with reef managers in French Polynesia, the first building blocks of an AI-based digital twin of the ocean.

MOFOMO (https://mofomo.inria.cl) takes over from OcéanIA, Inria’s research challenge on artificial intelligence for ocean science, led by Inria Chile for five years. In this new phase, Inria Chile will lead a multidisciplinary team spanning several institutions and three continents, with Inria project-teams in France and other scientific partners in Chile and French Polynesia.