Physics-based localization methodology for Data Assimilation by Ensemble Kalman Filter Journal: Journal of Computational Physics

Sarp ER, Marcello MELDI · 2025

Read Physics-based localization methodology for Data Assimilation by Ensemble Kalman Filter Journal: Journal of Computational Physics by Sarp ER, Marcello MELDI (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-08845.txt.

Introduction

Sarp Er and Marcello Meldi let the flow itself decide how an Ensemble Kalman Filter should localize its updates. Data assimilation blends simulations with measurements. Localization limits how far a measurement's influence reaches. It is usually drawn from fixed formulas that ignore what the fluid is actually doing.

The proposed localization function evolves with the instantaneous features of the ensemble members. It shrinks the regions that must be assimilated. That cuts the cost of coupling the filter to fluid dynamics solvers. Two test cases probe different physics: flow past a two-dimensional square cylinder at low Reynolds number, and turbulent flow around a three-dimensional circular cylinder at a Reynolds number of thirty-nine hundred. In both, accuracy holds while computational demands drop, which the authors call an excellent trade-off.

Catalog record

CREATORSarp ER, Marcello MELDI
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-08845

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-based localization methodology for Data Assimilation by Ensemble Kalman Filter Journal: Journal of Computational Physics Sarp ER, Marcello MELDI Abstract A physics-based methodology for the determination of the localization function for the Ensemble Kalman Filter (EnKF) is proposed. The spatial features of such function evolve dynamically over time according to the relevant instantaneous flow features of the ensemble members with the objective, to reduce the computational cost of the Data Assimilation (DA) procedure when applied with solvers for Computational Fluid Dynamics (CFD). The validation of the methodology has been carried out by the analysis of two test cases exhibiting different features. This permits to investigate different physical features, tailored for each test case, which affect the localization function. The flow over a two-dimensional square cylinder at R ​ e = 150 Re=150 is the first case investigated. It has been shown that the proposed localization procedure leads to a more cost-effective DA process by reducing the size of the assimilated regions while keeping the same level of accuracy. The capabilities of the methodology are further demonstrated by the investigation of the turbulent flow around a three-dimensional circular cylinder for R ​ e = 3900 Re=3900 . Again, the methodology exhibits an excellent trade off in terms of accuracy versus computational requirements. † † graphicalabstract: † † highlights: A physics based online-localization procedure is developed for the localization of the regions of data assimilation using EnKF. The proposed methodology is applied for the synchronization of the virtual model with a reference simulation in two different cases, namely 2D unsteady flow around square cylinder and 3D turbulent flow around circular cylinder. A significant reduction of the computational costs associated with DA procedure is observed due to the optimal choice of the DA regions at each analysis stage, without any loss in the synchronization performance as a result of the reduction of DA regions. The relationship between the optimized parameters and the flow physics have been demonstrated. 1 Introduction Applications of Data Assimilation (DA) [ 7 , 2 , 10 ] to fluid mechanics have seen a rapid raise in the last decade [ 31 , 26 , 19 , 17 , 37 , 5 , 20 , 8 , 39 , 33 , 32 , 22 ] . Thanks to the massive advancement in the availability of computational resources and in the development of efficient numerical architectures, one of the open challenges in the field is represented by the possibility to create digital twins [ 28 , 25 ] of complex flow applications using numerical simulations. In these applications, the numerical replica of the real flow would receive streaming information of the latter obtained at localized sensors and, thanks to its capacity to resolve large physical domain, would be able to produce control techniques to anticipate or delay unwanted physical occurrences. Because of the relevance of such applications in numerous fields such as urban and transport engineering, energy harvesting and pharmaceutical applications, research development for these technologies has seen a rapid increase in recent times. DA tools are perfectly tailored for this kind of tasks, as they are able to handle multiple streaming sources of information such as predictions obtained by the physical and the digital twin. Works available in the literature on this topic mainly deal with the usage of reduced order models (ROM) based on data-driven techniques [ 4 ] to perform the numerical tasks of a digital twin [ 9 ] . While the computational cost required by ROM is currently suitable for real-time investigation and coupling with physical phenomena, their performance over long time window may diverge due to intrinsic structural limitations and to quality of the data used for the training. The usage of high fidelity tools such as scale-resolving numerical simulation is still prohibitive for the application to digital twins for two main reasons. The first one is associated with the computational resources required to perform Direct Numerical Simulation (DNS) or Large Eddy Simulation (LES) [ 23 ] , which is prohibitively large at the present time. New paradigms in computational sciences, such as quantum computing, may however change the way calculations are performed in the foreseeable future. The second issue is associated with the computational costs of Data Assimilation to handle physical systems described by a large number of degrees of freedom (DOF). While DA has been successfully used to perform flow synchronization [ 33 ] and parametric optimization of the numerical model [ 34 , 33 , 32 ] , which are two cornerstones elements for efficient applications with digital twins, computational costs can be prohibitive if the DA algorithm is applied to the full physical domain. Some techniques of localization can be used to select specific regions of application of the DA procedures, potentially reducing the computational burdens required. Among these techniques, [ 35 ] have published a localization algorithm referred to as hyperlocalization . This strategy, which was developed for the Ensemble Kalman Filter (EnKF) [ 10 ] , a well known sequential DA algorithm, is able to reduce the computational requirement of more than 100 times with equivalent accuracy than classical DA approaches. The technique consists in applying the DA procedure only in selected region where important information is obtained by sensors. However, the size of this region is somehow arbitrary and the shape of the region has been selected to be spherical for practical reasons. In the present work, the hyperlocalization strategies previously introduced are augmented via infusion of the physical features of the flow. More precisely, the DA regions are shaped according to physical criteria related to the flow features observed in the region. The initial volume of such DA region, which is set to be spherical, is modulated to an ellipsoid whose axes and size are determined according to the physical criteria imposed. The new technique, which will be referred to as physics-based online localization (PBOL), is applied to investigate two unsteady test cases for external flows. The main difference between the two applications is the Reynolds number of investigation, which produces a laminar flow in the first case and a turbulent regime in the second application. Different physical criteria are applied to determine the size of the DA region for the two test cases, exploiting the features of the flow. The present article is structured as follows. In section 2 , the details of the numerical method for the solution of the governing equations are given, along with the ensemble Kalman filter algorithm used for the data assimilation stage and the various methods used for the localization of DA in the literature. In section 3 the novel approach for the localization, namely physics-based online-localization (PBOL), is introduced with a description of the hyperlocalization methodology. In section 4 , the results of the two test cases are presented. Firstly the 2D flow around the square cylinder, for the validation and the primary demonstration of the proposed PBOL methodology. Secondly the turbulent flow around 3D circular cylinder, where the application of the PBOL methodology is carried out for a fully turbulent test case. Finally, in section 5 , the conclusions and final remarks are given. 2 Numerical methods and algorithms This section is devoted to the presentation of the

Read the complete text