Building Tangent-Linear and Adjoint Models for Data Assimilation With Neural Networks

Article Meteorology & Atmospheric Sciences

A Hybrid Differential-Ensemble Linear Forecast Model for 4D-Var

T. J. Payne

Summary: The hybrid model proposed combines a simplified conventional tangent-linear model with LETLM-like adjustment for remaining processes, achieving better performance in tests compared to pure LETLM or existing linear models.

MONTHLY WEATHER REVIEW (2021)

添加到收藏夹

Article Multidisciplinary Sciences

Learning earth system models from observations: machine learning or data assimilation?

A. J. Geer

Summary: Recent progress in machine learning has inspired the idea of directly improving earth system models from observations, with data assimilation and machine learning sharing many similarities that can be unified under a Bayesian framework. Bayesian networks provide a practical framework for integrating data assimilation and machine learning to improve physical models of earth system processes.

PHILOSOPHICAL TRANSACTIONS OF THE ROYAL SOCIETY A-MATHEMATICAL PHYSICAL AND ENGINEERING SCIENCES (2021)

添加到收藏夹

Article Meteorology & Atmospheric Sciences

Machine Learning Emulation of Gravity Wave Drag in Numerical Weather Forecasting

Matthew Chantry et al.

Summary: Through machine learning training, we have developed emulators that produce more accurate results in long-term weather forecasting, with more complex networks producing more accurate emulators. In medium-range forecasting, our emulators are found to be more accurate than the parametrization scheme used for operational predictions.

JOURNAL OF ADVANCES IN MODELING EARTH SYSTEMS (2021)

添加到收藏夹

Article Meteorology & Atmospheric Sciences

Deep Emulators for Differentiation, Forecasting, and Parametrization in Earth Science Simulators

Marcel Nonnenmacher et al.

Summary: To overcome the limitation of missing derivatives in simulators, we propose using deep emulator networks that learn to calculate these derivatives. By training directly on simulation data without analyzing source code or equations, this approach supports simulators in any programming language on any hardware without specialized routines for each case. Demonstrating that emulator-derived derivatives enable accurate 4D-Var data assimilation and closed-loop training of parametrizations provides a basis for combining the parsimony and generality of physical models with the power and flexibility of machine learning.

JOURNAL OF ADVANCES IN MODELING EARTH SYSTEMS (2021)

添加到收藏夹

Article Meteorology & Atmospheric Sciences