Neural Data Assimilation for Regime Shift Monitoring of an Idealized AMOC Chaotic Model

Abstract

Data assimilation (DA) reconstructs and forecasts the dynamics of geophysical processes using available observations and physical a priori. Recently, the hybridization of DA and deep learning has opened new perspectives to address model-data interactions. This paper explores its potential contribution to the analysis of a chaotic oceanic phenomenon: the centennial to millennial variability of the North Atlantic ocean circulation during the last glacial period. The implemented neural approach—4DVarNet—yields meaningful improvements over a classical variational DA method in reconstructing regime shifts of the Atlantic Meridional Overturning Circulation (AMOC), especially when fewer observations are available. Interestingly, results exhibit that exploiting explicitly the a priori dynamical model does not lead to better performances compared to a data-driven model. Additionally, we compare four sampling strategies to assess how observation patterns influence the capture of unstable AMOC phases. We highlight the gain of regular over random sampling strategies, with reconstruction errors dropping below 2% for a 100-year sampling period. We find that monitoring the AMOC with regular clusters of three consecutive observation points can reduce errors by a factor of five. Eventually, we assess 4DVarNet's robustness in reconstructing a partially-observed system and in generalizing to different dynamical regimes. We also investigate on the maximum sampling periods that 4DVarNet can assimilate without compromising reconstruction quality. This study, based on an idealized yet complex physical model, suggests that neural approaches trained on observations wisely acquired could improve the monitoring of regime shifts in the context of climate change.