The objective is to study the acceleration of the equilibration of ocean currents in the IPSL global climate model (IPSLCM6). This equilibration phase also constrains the calibration of the parameters, and all of this ultimately represents 500% of the computational cost of the final simulations, used by scientific projects or by IPCC reports.
The acceleration by artificial intelligence models deployed on Jean Zay computer should greatly reduce this initialization phase of the climate models, and facilitate changes in spatial resolution in particular.
The idea is to have an AI inference model that extrapolates a series of simulation time steps (in months/years) and then reinjects the extrapolated solution into the climate model to perform new simulation steps. These two steps would be repeated as many times as necessary to obtain a stable algorithm that converges to a physically admissible solution comparable to full equilibria, while greatly reducing the number of time steps calculated explicitly with the model.
- an illustration of spin-up time from the QUEST project
Illustration of spin-up of IPSLCM6.2.2-MR025 (orange) and bifurcation towards another equilibrium (purple) after ocean currents were (mistakingly) reset to 0 ; other lines illustrate the long-term drift of IPSLCM6.2.2-MR1 (green and blue) and the bifurcation due to (intentional) change in CO2 concentration (yellow and red). Variable is the temperature at 1000m depth, globally averaged.
- illustration of the database used in the present project
hrmes project focuses on simulations of IPSLCM6 produced with NEMO v4.0 and v4.2. Because this version of NEMO is still relatively new, there are several simulations (of the coupled model) that only differ in a few ocean parameters. Hence their time evolution in ocean heat content is very close from one simulation to the other, yet the values reached after 250 years are significantly different from one another. In the figure below, the blue line represents a simulation that included a bug in the representation of freshwater fluxes, hence was not considered in our database. The other four simulations were included in the database used for training and testing.
In order to properly understand and accelerate the Spin-Up phase, we need to take into account the varying physics coming from the changing parameterization of our computational model(s) : the bigger and more varied our dataset, the better we should be able to extract a general behaviour, ensuring to some extent the viability of an acceleration method in all future simulations. On the other hand — the Spin-Up process potentially spanning time-scales up to the millenium — we need a dataset containing extensive lengths of simulation time, which goes against the need for a varied dataset since ressources are limited.
We tried to gather for a start 3 lengthy simulation runs, coming from previous versions of the NEMO model : this is what we defined as our Training Set (Sims 1-3). We then supplemented this first set with several shorter simulation runs using the current version of the NEMO model (v4.2.0–release candidate), as well as variations on the model parameters stemming from current development efforts : this is our Testing Set (Sims 4-7) on which acceleration attempts have been conducted.
↓↓↓ Table summing up the Available Simulation Dataset : 3 Training Simulations (1-3) and 4 Testing Simulations (4-7).↓↓↓
- The Barotropic Stream Function (MSFT)
As a starting point of the Spin-Up Acceleration Process, we studied the Barotropic Stream Function (also called Mass Stream FuncTion, MSFT).
↓↓↓ Time mean bartropic streamfunction (colours) and monthly anomalies in selected points (inserts) produced by Simulation #2↓↓↓
As we can see with the time series of the selected points, it is not straightforward to extract a trend reflecting the global adjustment of the barotropic streamfunction, which lasts months to decades depending on the region. This calls for statistical analysis of the dataset, to extract the underlying spin-up phase.
- Principal Component Analysis
For each simulation, low frequency variability in all modes but the first one, is very small. Hence the first mode can be considered as the underlying spin-up of the global barotropic streamfunction.
Applying PCA on the entire dataset or on individual simulations, yields very similar results, which suggests that what follows is not sensitive to changes in NEMO parameters.
Applying PCA on selected geographical regions does not yield the same result, ie some modes other than the first one, have significant variability at decadal to centennial time scales. This suggests that the result above is a property of the global barotropic streamfunction.
- Dynamic Mode Decomposition
to be completed - DMD has been tested but not yet conclusive as compared to PCA.
As long as the spin-up of the global barotropic streamfunction can be extracted with PCA, the complexity of our problem reduces to emulating the behaviour of a single time series.
We have explored several techniques such as :
- GRU
- ARIMA
- polynomial regression
- reservoir computing
- Gaussian processes
Inference by GP seems to be the most efficient technique :
We have made an attempt to restart the climate model using, for NEMO : altered velocities (which barotropic component was accelerated) but T, S, ssh and sea ice conditions unchanged.
The acceleration of the barotropic velocities was performed with GP over 10 years.
Restarting NEMO from hybrid conditions was much smoother than expected, ie no numerical instability appeared.
The resulting accelerated barotropic streamfunction behaves as expected : it is not the same as what would have happened with no acceleration, neither at time 0 nor 10 years later :
Other variables of the climate model, with accelerated barotropic velocities, behave as expected as well :
- those that have most variance at high frequency (sub-decadal) have very similar anomalies as ground truth (for example sea surface temperature),
- those that have most variance at lower frequency (decadal) show a discrepancy from ground truth, but similar anomalies at high frequency (for example temperature at 1000m depth).
There remain some inconsistencies, in particular in BGC and sea ice in northern hemisphere, to be clarified.
