Research Focus OCEAN

AI4PEX is focused on enhancing the understanding and modelling of ocean processes related to heat and carbon (C) uptake. The project aims to integrate machine learning (ML) with physical ocean models, creating a new generation of ML-enhanced coupled ESMs, to address key challenges in ocean science to improve the understanding of processes driving heat and carbon uptake in the ocean via:

1. Data Reconstructing: of ocean heat and carbon fluxes using ML to bridge scarce in situ with satellite-derived data for expanding data representativeness
2. Emulation: of ocean biogeochemistry models for expanding computational efficiency; and using deep probabilistic emulators to develop end-to-end neural schemes for model calibration and simulation of processes based on observation datasets.
3. Hybrid Modelling: formally bridging physical knowledge and ML in developing hybrid models, e.g. via neural mesoscale eddy parameterizations, to tackle major challenges in representing ocean dynamics. Ultimately, AI4PEX aims overcoming limitations of current methods and deliver richer representations of ocean heat and carbon uptake variabilities compared to current ensemble methods.

Here's a brief overview of planned tasks:

Task T1.2: Observations and Data Products in the Ocean

[lead: VLIZ; contributions: UHAM]

• Objective: Identify transport pathways of heat and carbon and improve the representation of marine heatwaves and their impacts.
• Approach: Upscale sparse in-situ ocean measurements and satellite data and use machine learning algorithms to harmonize global temperature and dissolved carbon fields as well as high-resolution coastal air-sea CO2 flux reconstructions.
• Focus Area:
  1. Meridional heat and carbon transport in the Southern Ocean.
  2. Air-sea CO2 exchange.

Task T2.3: ML-based Parametrizations for the Ocean

[lead: UREAD; contributions: MF-CNRM, IMT, UNIVLEEDS]

• Objective: Improve Earth System Model (ESM) parameterizations of processes that are key for oceanic heat and C uptake.
• Approach: Reduce model biases and increase computation efficiency using ML-based methods and apply neural closure methods to marine biogeochemical models.
• Focus Areas:
  1. Mesoscale eddy-induced transport.
  2. Marine biogeochemical cycles.

Task T3.3: Constraining and Quantifying Model Uncertainty for the Ocean

[lead: MF-CNRM; contributions: UREAD, UNIVLEEDS, IMT, CNRS-IGE]

• Objective: Quantify uncertainties in ocean heat and C uptake.
• Methods: Develop emulators for complex ocean biogeochemical models to better constrain uncertainties.
• Focus Area:
  1. Ocean heat and C uptake.
  2. Ocean biochemical response to physical forcings.

Task T4.3: Evaluation of ML-enhanced Ocean Models

[lead: UREAD; contributions: UNIVLEEDS, UVEG]

• Objective: Assess improvements in mesoscale eddy parameterization and their impact on ocean heat and C transport.
• Tools: Use explainable AI for sensitivity analysis and assess consistency in learned processes.
• Focus Area:
  1. Large-scale eddy-induced transports of heat and C,
  2. Surface eddy kinetic energy and relevant ocean currents.
  3. Overturning circulations and basin-scale ocean heat uptake.

Task T6.4: Coupling ML-enhanced Ocean and Sea-Ice Models

[lead: UNIVLEEDS; contributions: UREAD, CNRM]

• Objective: Evaluate key phenomena and trends in the new coupled and tuned Earth System Model (ESM).
• Approach: Integrate ML-enhanced ocean models with existing atmosphere, land, and sea-ice models.
• Focus Area:
  1. Radiative forcing and temperature.
  2. Surface and meridional fluxes of heat, water and C.
  3. Sequestration of heat and C to the deep ocean.

These tasks will ultimately advance ML-enhanced coupled Earth System Models, providing improved future projections of the Earth system.