Menu
MOFOMO home
White book PDF
  • EN
  • ES
  • FR
Spanish and French versions coming soon
Inria

0 mAbove the waterline

MOFOMO

Multimodal Foundation Models for a Digital Twin of the Ocean

Winner of Google.org Impact Challenge: AI for Science (2026 - 2029)

Nayat Sanchez-Pi and Luis Martí, Inria Chile Research Center

The ocean is the largest active sink of anthropogenic carbon, and the least observed part of the climate system.

MOFOMO learns a single model of it from the whole observational record, couples that model to the biology that moves the carbon, and tests it where the predictions have users. The approach, at 4,000 m

10 mThe problem

An incomplete and incoherent record

The ocean is the largest active sink of anthropogenic carbon and the least observed part of the climate system.

It covers 71% of Earth’s surface and absorbs about 90% of the excess heat and about a quarter of the CO₂ we emit: 3.4 ± 0.4 Pg C yr⁻¹ in 2024. That figure is well constrained. How fast the sink is changing is not.

Incomplete

Roughly 4,000 Argo floats, 1,000 biogeochemical floats, 1,000 moorings and 100 research vessels, against daily satellite coverage of the surface alone. Coverage is sparsest where it matters most: the Southern Ocean and the Global South.

Incoherent

Surface optics arrive daily at kilometre scale; interior profiles are sparse and irregular; imaging and genomics belong to a separate data culture. Each estimate is built within one modality, and results are compared only at the end.

Schematic ocean section: what the observing system sees A schematic cross-section of the ocean from the surface to 4,000 metres. Satellites cover the whole surface every day. Four Argo floats sample profiles down to 2,000 metres, one biogeochemical float, two moorings and one research-vessel cast reach deeper. Most of the interior between them is unobserved. 14 °C 10 °C 6 °C 4 °C 2 °C 0 200 500 1,000 2,000 3,000 4,000 m Satellite ocean colour: the whole surface, every day, surface only Argo float0–2,000 m, every 10 days Research vesselfull-depth cast Mooringfixed point Biogeochemical float unobserved unobserved
Schematic ocean section: what the observing system sees A schematic cross-section of the ocean from the surface to 4,000 metres. Satellites cover the whole surface every day. Argo floats sample profiles down to 2,000 metres; a biogeochemical float, a mooring and a research-vessel cast reach deeper. Most of the interior between them is unobserved. 14 °C 10 °C 6 °C 4 °C 2 °C 0 500 1,000 2,000 3,000 4,000 m Satellite: the whole surface, daily unobserved
What is observed. A schematic section from the surface to 4,000 m, with the depth axis stretched above 500 m. Satellites see the whole surface every day; below it, the ocean is seen only where an instrument happens to be. Counts are global.MOFOMO white book, §1.1
How fast did the ocean sink strengthen over 2001–2018? In Pg C yr⁻¹ per decade, three families of estimates disagree by a factor of about 1.8, on a quantity that enters every mitigation pathway.Gruber et al. (2023); DeVries et al. (2023), via the MOFOMO white book, §1.1

0–200 mSunlit zone

The record, depth by depth

Six kinds of observation will feed one model. Each sees its own slice of the water column, at its own resolution and rhythm. Follow them down.

  1. 0 msurface only

    Satellite ocean colour

    Sentinel-3 sees the colour of the sea every day, at kilometre scale. It is the only near-complete record of the ocean we have, and it stops at the surface.

    Sources
    Sentinel-3 ocean colour
    Rhythm
    Daily, km-scale
    Encoder
    Spherical / ViT
  2. 0–200 mwhere light reaches

    Plankton images

    Phytoplankton fix carbon in the sunlit zone and start the biological pump. Imaging surveys and occurrence records say who is there, and in what numbers.

    Sources
    OBIS, Planktonzilla, surveys
    Rhythm
    Cruises and time series
    Encoder
    Image (Planktonzilla CLIP init.)
    Plankton under the microscope: a copepod, chains of diatoms and round radiolarians on a pale background.
    Plankton under the microscope: the kind of image the image encoder learns from.

200–1,000 mTwilight zone

  1. 0–1,000 msampled by expedition, from the sunlit into the twilight zone

    Multi-omics

    Genomic samples describe what the community can do, not only who is in it. They come from expeditions, with their own resolutions and eras, and are rarely joined to the physical record.

    Sources
    Tara Oceans, BIO-GO-SHIP
    Rhythm
    Expeditions, 25+ years
    Encoder
    Genomic
  2. ~1,000 mbase of the twilight zone

    Passive acoustics

    Hydrophones record without pause, but each one is a single point. Sound carries far in the deep sound channel: whales, fish, rain and ships all leave a trace that a model can learn to read alongside the rest.

    Sources
    Moored hydrophones, observatories
    Rhythm
    Continuous, at fixed points
    Encoder
    Acoustic
    A school of silver fish swimming over a reef, shown in blue duotone.
    A multimodal model can ask how acoustic signatures track biodiversity.

1,000–2,000 mMidnight zone

  1. 0–2,000 mprofiled every 10 days

    Argo floats and reanalyses

    About 4,000 Argo floats drift at depth and profile the top 2,000 m every ten days, roughly 200 km apart. Reanalyses and climatologies fill the space between them, but they are weakly constrained this deep.

    Sources
    Argo, WOD, CMEMS, WOA23
    Rhythm
    ~200 km, 10 days
    Encoder
    Interpolation-free profile
    One Argo cycle: sink to 1,000 m, drift for about nine days, dive to 2,000 m, then rise to the surface while measuring, and transmit. 0 m1,0002,000 day 0day 10 drift at 1,000 m for ~9 days dive to 2,000 m measures on the way up transmit
    One float, one profile every ten days. The dotted leg is where it measures. Multiply by 4,000 and you have most of what we know about the ocean interior.

2,000–4,000 mMidnight zone, below the floats

What we don't observe

Down here the record runs out. What is left is sparse and uneven, and the estimates built on it do not agree.

The three families of methods used to estimate how fast the ocean carbon sink is strengthening disagree by a factor of about 1.8, on a number that enters every mitigation pathway assessment. The disagreement comes less from the physics of the models than from the record they are fitted to.

Bias in the estimated decadal trend of the ocean CO₂ flux, from sampling density alone % of the trend

Global ocean20–35 %
Southern Ocean50–130 %
Where the record is thinnest, the estimated trend is biased the most. The bias falls below 15 % only for an idealised network of about 1,000 sites.Hauck et al. (2023), via the MOFOMO white book, §1.1

Thinnest where it matters

Coverage is sparsest in the Southern Ocean and the Global South. Sampling density alone biases the estimated decadal trend of the flux by 20–35% globally, and by 50–130% in the Southern Ocean.

Compared too late

Each estimate is built within a single modality, and the results are compared only at the end, where disagreement is recorded rather than resolved.

Marine snow

The specks drifting down this page are marine snow. The biological pump exports on the order of 10 Pg C yr⁻¹ from the surface this way, and it is even less well constrained.

4,000 mAbyssal plain, where the dive turns

The approach: three linked blocks

The record cannot be enlarged. So MOFOMO builds one open system that learns across its gaps, couples what it learns to the biology, and tests both where predictions matter.

  • Block A

    Learning across the gaps

    A multimodal world model of the ocean: a foundation model pre-trained on the whole heterogeneous record. It is geometric (state on the sphere and on bathymetry-conforming meshes), latent-predictive, and read back out through symbolic regression and causal representation learning.

  • Block B

    Projecting the living ocean

    A coupled, trait-based biogeochemical model that goes beyond fixed plankton functional types (NPZ-T). It is written differentiably, so it trains together with block A.

  • Block C

    Validating where the impact is largest

    Two pilots where the carbon signal is large, the in-situ record dense and the stake immediate: the Patagonian inner sea of southern Chile and the reefs of Moorea, French Polynesia.

Modalities

Six streams, one latent ocean.

Each type of observation is encoded where, when and at the resolution it was measured, with no destructive regridding, into one structured latent space. Decoders then read out the physical, chemical and living ocean.

MOFOMO model: six modalities flow into one structured latent space and out to decoders and pilot modules Satellite ocean colour, in-situ profiles and reanalyses, passive acoustics, plankton images, multi-omics and scientific publications are each encoded at their native location, time and resolution into a shared latent space with modality-unique subspaces and a latent dynamics operator. Interventions (counterfactual forcings and hypothetical observations) condition the operator. Outputs are physical decoders, biogeochemical decoders, biodiversity heads, symbolic regression with causal discovery, the NPZ-T layer, and downstream modules for harmful algal bloom risk and reef bleaching. MODALITIES · NATIVE LOCATION, TIME, RESOLUTION STRUCTURED LATENT SPACE DECODERS, HEADS, MODULES Satellite ocean colourSentinel-3 · km-scale, daily In-situ profiles, reanalysesArgo · WOD · CMEMS · WOA23 Passive acousticscontinuous, point sensors Plankton imagesOBIS · Planktonzilla Multi-omicsTara Oceans · BIO-GO-SHIP Scientific publicationstext encoder Physical decodersT, S, currents Biogeochemical decoderssurface carbon state Biodiversity headscommunity structure Symbolic regression + causal discoveryclosed-form closures NPZ-T layerFloquet persistence thresholds Downstream modulesHAB risk · reef bleaching Shared subspace cross-modal alignment Modality-unique subspaces masked reconstruction Latent dynamics operator Interventions counterfactual forcings, hypothetical observations

Modalities · native location, time, resolution

  • Satellite ocean colourSentinel-3 · km-scale, daily
  • In-situ profiles, reanalysesArgo · WOD · CMEMS · WOA23
  • Passive acousticscontinuous, point sensors
  • Plankton imagesOBIS · Planktonzilla
  • Multi-omicsTara Oceans · BIO-GO-SHIP
  • Scientific publicationstext encoder

Structured latent space

Shared subspacecross-modal alignment
Modality-unique subspacesmasked reconstruction
Latent dynamics operator
conditions the operatorInterventionscounterfactual forcings, hypothetical observations

Decoders, heads, modules

  • Physical decodersT, S, currents
  • Biogeochemical decoderssurface carbon state
  • Biodiversity headscommunity structure
  • Symbolic regression + causal discoveryclosed-form closures
  • NPZ-T layerFloquet persistence thresholds
  • Downstream modulesHAB risk · reef bleaching
  • Pre-training is contrastive, then reconstructive, with modality dropout, so the model learns what the streams share and what only one of them sees.
  • Conditioning is score-based: the operator accepts any set of observations, including hypothetical ones, without retraining.
  • Every output is calibrated: deep ensembles and conformal prediction wrap each prediction with its uncertainty.
  • NPZ-T is a nutrient-phytoplankton-zooplankton model with plankton traits; its Floquet thresholds tell whether a population can persist through the seasonal cycle.

2,000 m ↑On the way back up

Five research questions

What the project has to answer, in short. The white book sets each one out in full, with the evidence that would refute it.

  • Representation

    Can one foundation model learn unified representations across satellite imagery, in-situ profiles, acoustic signals, and biological and genomic observations that outperform single-modality specialists?

  • Geometry

    Do spherical and mesh operators, frame-independent vector fields and flux-form tendencies improve data efficiency, physical plausibility and transfer, against a gridded transformer of matched capacity?

  • World model

    Can the model predict ocean and ecosystem states under counterfactual forcings and hypothetical observations, rather than only nowcast the trajectory we observed?

  • Mechanism

    Can symbolic regression on the learned operator, paired with causal discovery, recover closed-form relations stable enough to insert into a trait-based model?

  • Transfer

    Does cross-modal pre-training transfer to biogeochemical and biodiversity tasks where labelled data is scarce?

200 m ↑Back towards the coast

Pilot studies

Two contrasting coastal systems where the carbon signal is large, the in-situ record dense and the stake immediate. In both, predictions can be checked at the scale where decisions are made.

Plate IChiloé–Aysén · 42–46° S · 73° W

The Patagonian inner sea of southern Chile

Plate I. Patagonian inner sea, Chiloé to AysénSchematic bathymetric map of the inner sea of southern Chile from Puerto Montt (41.5° S) to Puerto Aysén (45.4° S), showing Chiloé Island, the Golfo de Ancud and Golfo Corcovado, the Chonos Archipelago, the main rivers, isobaths from 100 to 3000 metres and the open boundary at the Boca del Guafo where the shelf–fjord exchange flux is scored. 200 500 2,000 Open boundary: shelf–fjord flux scored here Puerto Montt Castro Chaitén Melinka Puerto Aysén Pacific Ocean Golfo de Ancud Golfo Corcovado Canal Moraleda Chiloé Chonos Archipelago LOS LAGOS AYSÉN Palena 42° S 43° S 44° S 45° S 75° W 74° W 73° W 72° W 0 50 100 km N
0 m1002001,000+ open boundary rivers isobaths 100–3,000 m Schematic; not for navigation.

A fjord–estuarine system driven by river and glacial discharge, with intensive salmon aquaculture and hypoxic deep basins. Harmful algal blooms here are episodic reorganisations of the coastal carbon cycle.

Asks
How would a bloom have developed had river discharge held up, or the shelf upwelling relaxed, two weeks earlier?
Delivers
Harmful-algal-bloom risk advisories for salmon farming
Target
AUROC ≥ 0.80
Mesh
Refined to kilometre scale; the shelf–fjord exchange across the open boundary is a scored flux
Diatoms under the microscope, glassy cells in blue light.
Diatoms: some bloom-forming species are harmful to farmed fish.
Plate IIMoorea · 17.5° S · 149.8° W

South Pacific island reefs: Moorea, French Polynesia

Plate II. Moorea, French PolynesiaSchematic bathymetric map of Moorea and the western coast of Tahiti, showing the lagoon and reef crest, Opunohu and Cook’s bays, the Gump South Pacific Research Station and CRIOBE, and isobaths from 100 to 3000 metres. 500 2,000 3,000 Gump Station CRIOBE Moorea Tahiti Pacific Ocean reef crest 17.4° S 17.5° S 17.6° S 17.7° S 149.9° W 149.8° W 149.7° W 149.6° W 0 5 10 km N
0 m1005001,000+ reef crest field stations Schematic; not for navigation.

A remote island reef with resident field teams, where bleaching shifts the reef’s carbonate balance from net precipitation towards net dissolution. With the Gump South Pacific Research Station (UC Berkeley) and CRIOBE.

Asks
Would this reef have bleached under the previous decade’s thermal history?
Delivers
Coral-bleaching early warning for reef managers
Baseline
Degree Heating Week index, AUROC 0.758
Data
Moorea Coral Reef LTER, CRIOBE observatory records
A blue sea star among branching and massive corals.
A reef in colour: near the surface, the light comes back.

10 m ↑Open science

Everything comes back up, in the open

Every part of the system is released openly, so that others can check it, reuse it and build on it.

External validation

Release manifest
WhatLicenceWhere
Foundation-model backboneApache 2.0 / OpenRAILHugging Face
Modality encodersApache 2.0 / OpenRAILHugging Face
Latent dynamics operatorApache 2.0 / OpenRAILHugging Face
Recovered closed-form expressionsCC-BY 4.0Papers, GitHub
Multimodal data lakeCC-BY 4.0 / CC0Hugging Face
Benchmark and public leaderboardApache 2.0 · CC-BY 4.0GitHub
Downstream modules: HAB, bleachingApache 2.0 / OpenRAILHugging Face

5 m ↑Publications

Publications and talks

MOFOMO's publications along with the papers, talks, theses and reports of OcéanIA.

In the bibliography

1
MOFOMO entry, 2026
50
OcéanIA entries, 2020–2026

All publications

MOFOMO latest

  1. Book2026

    Working document, version 0.1: Sanchez-Pi, N., & Martí, L. (2026). MOFOMO: Multimodal Ocean Foundation Models for an Ocean Digital Twin. Santiago, Saclay, Sophia Antipolis, Paris, Moorea: Inria – Institut national de recherche en sciences et technologies du numérique.

OcéanIA latest

  1. Report2026

    Contreras Montanares, A. G., Valenzuela, L., Martí, L., & Sanchez-Pi, N. (2026). Planktonzilla: Multimodal dataset and models for understanding plankton ecosystems. Inria Chile Research Center.

  2. Conference paper2026

    Valenzuela, L., Aguilera Madrid, S., Martí, L., & Sanchez-Pi, N. (2026). Benchmarking COX1 Embeddings across Genomic Foundation Models, kmers, and Imbalance-Aware Losses. In Fourteenth International Conference on Learning Representations (ICLR 2026) Workshop on Generative AI in Genomics (Gen2): Barriers and Frontiers.

  3. Conference paper2026

    Valenzuela, L., Aguilera Madrid, S., Martí, L., & Sanchez-Pi, N. (2026). From k-mers to Genomic Foundation Models: Benchmarking COX1 Taxonomy under Extreme Class Imbalance. In Fourteenth International Conference on Learning Representations (ICLR 2026) Workshop on on Machine Learning for Genomics Explorations.

All OcéanIA publications

0 mOn board

From OcéanIA to MOFOMO

OcéanIA logoOcéanIA, Inria Challenge
2020–2025

OcéanIA brought AI and mathematical modelling to the ocean for five years. MOFOMO extends its results into one foundation model.

  1. OcéanIA starts as an Inria Challenge, led from Inria Chile.

  2. First AIMOCC workshop, at ICLR. OcéanIA white book published.

  3. AIMOCC at IJCAI, and the OcéanIA challenge at IJCAI.

  4. AIMOCC at ECAI.

  5. OcéanIA selected by the Paris Peace Forum among 50 public-interest AI projects.

  6. OcéanIA presented at the AI Action Summit, Paris.

  7. OcéanIA at the UN Ocean Conference (UNOC3), Nice.

  8. MOFOMO begins with support of the Google.org Impact Challenge: AI for Science (2026 – 2029).

Partners

  • Inria ChileHost and project lead, Santiago, Chile
  • Inria TAU teamMachine-learning foundations, large-scale training, Paris, France
  • Inria GREENOWL teamPhytoplankton and microbial-ecosystem modelling, Sophia Antipolis, France

Note: We are finalizing the list of Project Partners.

Team

  • Nayat SÁNCHEZ-PIPI - Director, Inria Chile
  • Luis MARTÍCo-PI - Scientific Director, Inria Chile
  • Michèle SEBAGTAU project team; Inria Paris-Saclay Centre
  • Marc SCHOENAUERTAU project team; Inria Paris-Saclay Centre
  • Sylvain CHEVALIERTAU project team; Inria Paris-Saclay Centre
  • Francesca CASAGLIBIOCORE project team - Inria Centre at Université Côte d'Azur
  • Olivier BERNARDBIOCORE project team; Inria Centre at Université Côte d'Azur
  • Luis VALENZUELAInria Chile
  • Hernán LIRAInria Chile
  • Matías NETOInria Chile
  • Funding

    Inria Chile receives US$2.5 million from Google.Org for MOFOMO

    Inria Chile is one of 17 recipients of the Google.org Impact Challenge: AI for Science, chosen from more than 3,000 applicants, and the only one from Latin America. The grant funds MOFOMO for 36 months.

  • White book

    MOFOMO white book v0.1 released

    The first working version sets out the research programme: the incomplete record, the three-block system, the research questions and how success will be measured.