[Paper Review] DAN - An optimal Data Assimilation framework based on machine learning Recurrent Networks.

Pierre Boudier, Anthony Fillion+2 more|arXiv (Cornell University)|Oct 19, 2020

Meteorological Phenomena and Simulations21 references4 citations

TL;DR

This paper proposes DAN, a deep learning framework that combines recurrent neural networks with data assimilation principles to optimally reconstruct and forecast the state of dynamical systems from noisy observations. Trained on trajectory snapshots, DAN accurately models both the analysis and propagation of probability density functions, matching the theoretical performance of traditional data assimilation methods while being fully data-driven.

ABSTRACT

Data assimilation algorithms aim at forecasting the state of a dynamical system by combining a mathematical representation of the system with noisy observations thereof. We propose a fully data driven deep learning architecture generalizing recurrent Elman networks and data assimilation algorithms which provably reaches the same prediction goals as the latter. On numerical experiments based on the well-known Lorenz system and when suitably trained using snapshots of the system trajectory (i.e. batches of state trajectories) and observations, our architecture successfully reconstructs both the analysis and the propagation of probability density functions of the system state at a given time conditioned to past observations.

Motivation & Objective

• To develop a fully data-driven deep learning architecture that generalizes both recurrent networks and data assimilation algorithms.
• To achieve the same prediction accuracy as traditional data assimilation methods without relying on explicit system models.
• To reconstruct the probability density functions of system states conditioned on past observations through end-to-end training.
• To demonstrate the framework's ability to model both analysis and propagation of state uncertainties in nonlinear dynamical systems.

Proposed method

• The framework is based on a modified recurrent Elman network architecture adapted to handle sequential observations and state trajectories.
• It uses batches of state trajectories as training data to learn the mapping from observations to posterior state distributions.
• The network is trained to minimize a loss function that enforces consistency with the principles of optimal data assimilation.
• The architecture incorporates memory mechanisms to propagate uncertainty information across time steps, mimicking the analysis and forecast cycles of data assimilation.
• The training process leverages temporal dependencies in trajectory data to learn the system's dynamics and observation noise characteristics.
• The method is theoretically grounded in proving that the network's predictions converge to the same optimal solutions as classical data assimilation algorithms.

Experimental results

Research questions

• RQ1 Can a deep learning architecture be designed to replicate the optimal performance of traditional data assimilation methods in state estimation?
• RQ2 How well can a recurrent neural network trained on trajectory snapshots reconstruct the full probability density function of a system's state conditioned on past observations?
• RQ3 To what extent can a purely data-driven approach outperform or match model-based data assimilation in nonlinear dynamical systems?
• RQ4 What architectural modifications are necessary to ensure that a recurrent network learns the correct analysis and forecast steps of data assimilation?
• RQ5 Does the network's ability to model uncertainty propagation match theoretical expectations from optimal estimation theory?

Key findings

• The DAN framework successfully reconstructs both the analysis and propagation of probability density functions of the system state, conditioned on past observations.
• The network achieves performance comparable to optimal data assimilation algorithms when trained on trajectory snapshots and observational data.
• The architecture generalizes recurrent Elman networks by embedding the mathematical structure of data assimilation into its learning process.
• The framework demonstrates robustness in modeling nonlinear dynamics, as validated on the Lorenz system.
• Theoretical analysis confirms that the network's predictions converge to the same optimal solutions as classical data assimilation methods.
• The method enables end-to-end learning of state estimation and uncertainty propagation without requiring explicit knowledge of the system's governing equations.