Data assimilation (DA) is integrated with machine learning in order to perform entirely data‐driven online state estimation. To achieve this, recurrent neural networks (RNNs) are implemented as pretrained surrogate models to replace key components of the DA cycle in numerical weather prediction (NWP), including the conventional numerical forecast model, the forecast error covariance matrix, and the tangent linear and adjoint models. It is shown how these RNNs can be initialized using DA methods to directly update the hidden/reservoir state with observations of the target system. The results indicate that these techniques can be applied to estimate the state of a system for the repeated initialization of short‐term forecasts, even in the absence of a traditional numerical forecast model. Further, it is demonstrated how these integrated RNN‐DA methods can scale to higher dimensions by applying domain localization and parallelization, providing a path for practical applications in NWP.
12 May 2021 in Journal of Advances in Modeling Earth Systems
A description and assessment of the first release of the Arctic Subpolar gyre sTate Estimate (ASTE_R1), a data‐constrained ocean‐sea ice model‐data synthesis, is presented. ASTE_R1 has a nominal resolution of 1/3° and spans the period 2002–2017. The fit of the model to an extensive (O(109)) set of satellite and in situ observations was achieved through adjoint‐based nonlinear least squares optimization. The improvement of the solution compared to an unconstrained simulation is reflected in misfit reductions of 77% for Argo, 50% for satellite sea surface height, 58% for the Fram Strait mooring, 65% for Ice Tethered Profilers, and 83% for sea ice extent. Exact dynamical and kinematic consistency is a key advantage of ASTE_R1, distinguishing the state estimate from existing ocean reanalyses. Through strict adherence to conservation laws, all sources and sinks within ASTE_R1 can be accounted for, permitting meaningful analysis of closed budgets at the grid‐scale, such as contributions of horizontal and vertical convergence to the tendencies of heat and salt. ASTE_R1 thus serves as the biggest effort undertaken to date of producing a specialized Arctic ocean‐ice estimate over the 21st century. Transports of volume, heat, and freshwater are consistent with published observation‐based estimates across important Arctic Mediterranean gateways. Interannual variability and low frequency trends of freshwater and heat content are well represented in the Barents Sea, western Arctic halocline, and east subpolar North Atlantic. Systematic biases remain in ASTE_R1, including a warm bias in the Atlantic Water layer in the Arctic and deficient freshwater inputs from rivers and Greenland discharge.
Drawing on ergodic theory, we introduce a novel training method for machine learning based forecasting methods for chaotic dynamical systems. The training enforces dynamical invariants—such as the Lyapunov exponent spectrum and the fractal dimension—in the systems of interest, enabling longer and more stable forecasts when operating with limited data. The technique is demonstrated in detail using reservoir computing, a specific kind of recurrent neural network. Results are given for the Lorenz 1996 chaotic dynamical system and a spectral quasi-geostrophic model of the atmosphere, both typical test cases for numerical weather prediction.
21 November 2020 in Journal of Geophysical Research: Oceans
Ocean bathymetry exerts a strong control on ice sheet‐ocean interactions within Antarctic ice‐shelf cavities, where it can limit the access of warm, dense water at depth to the underside of floating ice shelves. However, ocean bathymetry is challenging to measure within or close to ice‐shelf cavities. It remains unclear how uncertainty in existing bathymetry datasets affect simulated sub‐ice‐shelf melt rates. Here we infer linear sensitivities of ice‐shelf melt rates to bathymetric shape with grid‐scale detail by means of the adjoint of an ocean general circulation model. Both idealized and realistic‐geometry experiments of sub‐ice‐shelf cavities in West Antarctica reveal that bathymetry has a strong impact on melt in localized regions such as topographic obstacles to flow. Moreover, response of melt to bathymetric perturbation is found to be non‐monotonic, with deepening leading to either increased or decreased melt depending on location. Our computational approach provides a comprehensive way of identifying regions, where refined knowledge of bathymetry is most impactful, and also where bathymetric errors have relatively little effect on modeled ice sheet‐ocean interactions.
There has been a recent surge in development of accurate machine learning (ML) weather prediction models, but evaluation of these models has mainly been focused on medium‐range forecasts, not their performance in cycling data assimilation (DA) systems. Cycling DA provides a statistically optimal estimate of the system state, which can then be used as initial conditions for model prediction, given observations and previous model forecasts. Here, real surface pressure observations are assimilated into several popular ML models using an ensemble Kalman filter, where accurate ensemble covariance estimation is essential to constrain unobserved state variables from sparse observations. In this cycling DA system, deterministic ML models accumulate small‐scale noise until they diverge. Mitigating this noise with a spectral filter can stabilize the system, but with larger errors than traditional models. Perturbation experiments illustrate that these models do not accurately represent short‐term error growth, leading to poor estimation of cross‐variable covariances.