Abstract : Weather is a good example of large-scale chaotic complex systems, with a strong sensitivity to initial conditions tied to the intrinsic limit to predictability. The sensitivity suggests effective control in which small modifications to the atmospheric conditions grow rapidly and result in big changes. Weather predictability has been studied extensively in the past decades, and prediction skills have been improving. With accurate weather prediction, we are now ready to study weather controllability. This mini-symposium consists of solicited presentations about the mathematics behind the predictability and controllability of complex systems such as weather and other problems.
Organizer(s) : Takemasa Miyoshi, Sebastian Reich, Takashi Sakajo, Kohei Takatama
[02251] Sequential data assimilation and data driven control
Format : Online Talk on Zoom
Author(s) :
Sebastian Reich (University of Potsdam)
Abstract : Sequential data assimilation can be considered a coupling of measure problem which can be tackled using optimal transport or Schroedinger bridges. In this talk, we will present an approximate bridging approach which leads to a data-driven term being added to the underlying model dynamics. The added control term is of mean-field type and can be implemented easily within a Monte Carlo context. The control term is also closely related to continuous time ensemble Kalman-Bucy filter formulations. One can extend the proposed controlled DA approach to cost functions other than those derived from the data likelihood.
[02790] Data-driven Reconstruction of Partially Observed Dynamical Systems
Format : Talk at Waseda University
Author(s) :
Pierre Tandeo (IMT Atlantique)
Pierre Ailliot (Univ. Brest)
Florian Sévellec (CNRS)
Abstract : The goal of this work is to obtain predictions of a partially observed dynamical system, without knowing the model equations. To account to those strong assumptions, a combination of machine learning and data assimilation techniques is proposed with the introduction of latent variables. We find that the latent variables inferred by the procedure are related to the successive derivatives of the observed components of the dynamical system.
[02800] Sensor selection by greedy method for linear dynamical systems
Format : Talk at Waseda University
Author(s) :
Shun Takahashi (Tokai University)
Kumi Nakai (National Institute of Advanced Industrial Science and Technology )
Takayuki Nagata (Tohoku University)
Keigo Yamada (Tohoku University)
Yasuo Sasaki (Tohoku University)
Yuji Saito (Tohoku University)
Taku Nonomura (Tohoku University)
Abstract : Sensor optimization using a greedy method based on the snapshot-to-snapshot Fisher information matrix, observability Gramian, and Kalman filter indices in linear time-invariant systems is discussed. The objective functions and computational times are compared for the resulting sensor sets with a background of application to sensor selection in large systems.
[02804] Fast Linear-regression-based Sensor Selection and its Applications
Format : Talk at Waseda University
Author(s) :
Yasuo Sasaki (Tohoku University)
Yuji Saito (Tohoku University)
Takayuki Nagata (Tohoku University)
Keigo Yamada (Tohoku University)
Taku Nonomura (Tohoku University)
Abstract : We consider a ridge-regression-based sensor selection problem in which sensors are selected so that dependent variables can be estimated as easily as possible. For this problem, a fast greedy algorithm is derived by means of one-rank update law of covariance matrices. To verify the effectiveness of this greedy algorithm, it is applied to sensor selection for estimation of the sea surface temperature and for optimal feedback control of flow around a circular cylinder.
[02789] Identifying Coherent Structures within Turbulent Flows over Roughness Obstacles
Format : Talk at Waseda University
Author(s) :
Tetsuya Takemi (Kyoto University)
Abstract : Large-eddy simulations for turbulent flows over roughness obstacles are conducted to investigate the characteristics of coherent structures within turbulent flows for a possible application for controlling flow structures in urban districts. Idealized and realistic urban districts are examined as rough surfaces. Turbulence and dispersion properties associated with coherent structures are demonstrated. Understanding such properties will lead to identify an approach to control flows within urban districts.
[02825] Quantifying Weather Controllability and Mitigatable Flood Damage Based on Ensemble Weather Forecast
Format : Talk at Waseda University
Author(s) :
Shunji Kotsuki (Chiba University)
Abstract : For realizing a weather-controlled society, we need to discuss the way to maximize the effect of manipulations to the atmosphere. For that purpose, this project aims at developing methods that quantify weather controllability and mitigatable flood damage based on ensemble weather forecasts. To quantify weather controllability, this project investigates meteorological landscapes that separate disaster and non-disaster regimes which may be controllable with small manipulations. We also estimate economic damages under non-controlled/controlled scenarios, in order to quantify avoidable damage by weather control.
We have started illustrating directed graphs as the first step in understanding the meteorological landscape. Typhoon Prapiroon in 2018 was used for the case study. Singular value decomposition (SVD) is employed for Japan Meteorological Agency’s operational meso-scale ensemble prediction data to extract principle components of atmospheric states, followed by a clustering using density-based spatial clustering of applications with noise known as DBSCAN. The illustrated graph succeeded in detecting separated two clusters that correspond to faster and slower movements of predicted Parapiroon. The developed algorithm is currently applied to other disastrous events as well as further investigations on non-linear data compression methods beyond SVD. This presentation includes the most recent achievements up to the time of the conference.
[02885] Ensemble sensitivity and its potential applications in weather control
Author(s) :
Le Duc (University of Tokyo)
Yohei Sawada (University of Tokyo)
Abstract : Ensemble sensitivity has been proved to be a very useful sensitivity measure in practice. In this study, we show the relevance of ensemble sensitivity in another important problem. We have proved that changes of the forecast response are maximized along the direction of the vector consisting of ensemble sensitivities which forms the most sensitive perturbation. We will demonstrate how the new understanding on ensemble sensitivities can qualitatively give potential solutions for weather control.
[02518] Chaos implies effective controllability of extreme weather
Format : Talk at Waseda University
Author(s) :
Takemasa Miyoshi (RIKEN)
Qiwen Sun (RIKEN)
Koji Terasaki (RIKEN)
Yasumitsu Maejima (RIKEN)
Abstract : Since the weather system is chaotic, and even more so for storms, small differences generally lead to big differences, particularly for high-impact weather events. This presentation will summarize the concept and methodology of Control Simulation Experiment (CSE) with some proof-of-concept demonstrations with toy models and realistic Numerical Weather Prediction (NWP) models. This is an attempt to a potential paradigm change of NWP research from decades of predictability to the new era of controllability.
[02814] Noise Calibration for the Stochastic Rotating Shallow Water Equations
Format : Talk at Waseda University
Author(s) :
Alexander Lobbe (Imperial College London)
Oana Lang (Imperial College London)
Dan Crisan (Imperial College London)
Peter Jan van Leeuwen (Colorado State University)
Roland Potthast (Deutscher Wetterdienst DWD)
Abstract : We introduce a new method of noise calibration of the Stochastic Rotating Shallow Water (SRSW) model which is rigorously derived from a model reduction technique. The method is generic and can be applied to arbitrary stochastic models. In the (SRSW) case, we calibrate the noise by using the pressure variable of the model, as this is an observable easily obtainable in practical application.
[03049] Machine learning-based estimation of state-dependent forecast uncertainty
Format : Talk at Waseda University
Author(s) :
Juan Jose Ruiz (University of Buenos Aires)
Maximiliano Sacco (National Meteorological Service of Argentina)
Manuel Pulido (Universidad Nacional del Nordeste)
Pierre Tandeo (IMT Atlantique)
Abstract : Quantifying forecast uncertainty is a key aspect of state-of-the-art numerical weather prediction and data assimilation systems. State dependent uncertainty quantification in numerical weather prediction is a computation intensive task which has been performed using different approaches such as monte carlo sampling (ej. ensemble Kalman filter) and variational approaches (ej. adjoint based model sensitivity). Machine learning techniques consist of trainable statistical models that can represent complex functional dependencies among different groups of variables given a large enough dataset. In this talk we will describe the use of a machine learning approach based on neural networks for the estimation of forecast uncertainty. In particular, we will discuss the estimation of the forecast error covariance matrix, which is at the center of probabilistic forecasting and data assimilation systems. In addition, we will present a hybrid data assimilation method that combines the optimal interpolation technique and a convolutional neural network to estimate the state dependent forecast error covariance matrix.
[03126] Observability of continuous-time Markov model and filter stability
Format : Talk at Waseda University
Author(s) :
JIN WON KIM (University of Potsdam)
Abstract : In control theory, estimation and control are considered as dual problems. A fundamental relationship is the duality between controllability and observability, and it extends to Kalman filter and a linear quadratic control problem. Our contribution is to extend the duality to nonlinear models. I will review the classical duality and present the dual optimal control problem. The dual formulation is used to analyze the stability of the nonlinear filter, similar to the linear Gaussian case.
[03024] On random feature maps in prediction
Format : Talk at Waseda University
Author(s) :
Nicholas Cranch (University of Sydney)
Georg A. Gottwald (University of Sydney)
Sebastian Reich (University of Potsdam)
Abstract : Random feature maps (RFs) can be viewed as a single hidden layer network in which the weights of the hidden layer are fixed. We show how the choice of the internal weights effects performance and generalisation. We propose how to best choose the internal weights. We show that RFs allow for sequential learning when combined with data assimilation, and can be used to learn subgridscale parametrizations and to detect critical transitions.
