Developing a data and physics driven machine learning mesoscale eddy parameterisation for NEMO

The eddy-permitting NEMO model (ORCA025) is known to exhibit sub-par Southern Ocean circulation features, such as a too weak Antarctic Circumpolar Current transport and cool and warm biases on the Antarctic shelf. The ORCA025 model sits in the numerical grey zone, which is where the horizontal grid resolution can only resolve mesoscale processes over part of the domain. In other parts of the domain, the eddies need to be parameterised, such as high-latitude regions. This difficulty in representing eddies has in-part contributed to the poor Southern Ocean circulation, leading to great uncertainty in key climate metrics such as carbon and heat transport, and the Antarctic ice mass balance. The key question is, how do we parameterise mesoscale eddies where they are most needed, without being detrimental to the resolved flow. Scale- and flow-aware parameterisations have been implemented in NEMO and have led to improvements in some flow characteristics. However, an alternative approach is to leverage data, physics, and machine learning to develop an improved eddy parameterisation.

As part of the project, AI4PEX, we aim to develop a data- and physics-driven mesoscale eddy parameterisation that better captures the dynamical feedback between mesoscale eddies and the large-scale ocean circulation, reducing model uncertainty. In our work, we will attempt to improve an eddy parameterisation that is available in NEMO, GEOMETRIC. To do this we will use a Neural Network trained on high resolution data from realistic global models ORCA12/ORCA36. To reduce the black-box nature of the Neural Network, we will design a loss function that is informed by the physics of mesoscale eddies. Initial investigation of the eddy parameterisation will take place offline in an idealised configuration.