Description
During my end-of-study internship at the LOCEAN-IPSL laboratory, I have been working on coupled ocean-atmosphere models calibration.
The ccentral goal of the the project was to better understand if History Matching technique was adapated for calibrating coupled ocean-atmosphere models. Using a simplified model, the two layers Lorenz-96 we demonstrated that History Matching method effectively constrained the model’s parameter search space. Notably, this was validated across scenarios resembling AMIP (Atmospheric Model Intercomparison Project) and OMIP (Ocean Model Intercomparison Project) experiments.
We also experimented the use of machine learning alternative to gaussian processes as emulators. We carried a comparative study of Ramdom Forest and Bayesian Neural Networks with the commonly used gaussian process. While these methods do not seem to show strong performance improvement in the convergence speed to the real parameters, they seem to perform well and some further investigation might be promising.
We finally compared two approaches to reduce the metric space: Principal Component Analysis (as this is usually used) and autoencoders. We show that by retaining more information about the metrics in a smaller subspace, the former improve the algorithm convergence speed and accuracy.