Toward data-driven methods in geophysics: the Analog Data Assimilation.

LGUENSAT R., P. TANDEO, P. AILLIOT, M. PULIDO, AND R. FABLET

Vienna

Conferencia; European Geosciences Union General Assembly; 2017

European Geosciences Union

The Analog Data Assimilation (AnDA) is a recently introduced data-drivenmethods for data assimilation where the dynamical model is learned fromdata, contrary to classical data assimilation where a physical model ofthe dynamics is needed. AnDA relies on replacing the physical dynamicalmodel by a statistical emulator of the dynamics using analog forecastingmethods. Then, the analog dynamical model is incorporated inensemble-based data assimilation algorithms (Ensemble Kalman Filter andSmoother or Particle Filter). The relevance of the proposed AnDA isdemonstrated for Lorenz-63 and Lorenz-96 chaotic dynamics. Applicationsin meteorology and oceanography as well as potential perspectives thatare worthy of investigation are further discussed. We expect that thedirections of research we suggest will help in bringing more interest inapplied machine learning to geophysical sciences.