A Comparison of Parametric Dynamic Mode Decomposition Algorithms for Thermal-Hydraulics Applications

Stefano Riva, Andrea Missaglia, Carolina Introini, In Cheol Bang, Antonio Cammi · 2025

Read A Comparison of Parametric Dynamic Mode Decomposition Algorithms for Thermal-Hydraulics Applications by Stefano Riva, Andrea Missaglia, Carolina Introini, In Cheol Bang, Antonio Cammi (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-2503-24205.txt.

Introduction

Learning a system's behavior from data can shrink heavy simulations into light surrogate models. Stefano Riva and co-authors compare in 2025 the family of parametric dynamic mode decomposition algorithms. Such methods extract a linear stand-in for a dynamical system. They must also handle changing parameters, not just one run. Three thermal-hydraulics problems serve as the arena.

Two are cylinder-flow benchmarks, one computed fresh, one public. The third is a real experimental loop studying natural circulation for future reactors. The survey weighs each variant's strengths and failures, giving modelers a map of which decomposition to trust for parameter-swept predictions.

Catalog record

CREATORStefano Riva, Andrea Missaglia, Carolina Introini, In Cheol Bang, Antonio Cammi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE74 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-24205

Browse the full library · About the Tower of Babel

Full text

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

A Comparison of Parametric Dynamic Mode Decomposition Algorithms for Thermal-Hydraulics Applications Stefano Riva, Andrea Missaglia, Carolina Introini, In Cheol Bang, Antonio Cammi Abstract In recent years, algorithms aiming at learning models from available data have become quite popular due to two factors: 1) the significant developments in Artificial Intelligence techniques and 2) the availability of large amounts of data. Nevertheless, this topic has already been addressed by methodologies belonging to the Reduced Order Modelling framework, of which perhaps the most famous equation-free technique is Dynamic Mode Decomposition. This algorithm aims to learn the best linear model that represents the physical phenomena described by a time series dataset: its output is a best-fit linear operator of the underlying dynamical system that can be used, in principle, to advance the system itself in time even beyond the training time interval. However, in its standard formulation, this technique cannot deal with parametric time series, meaning that a different linear model has to be derived for each configuration. Research on this is ongoing, and some versions of a parametric Dynamic Mode Decomposition already exist. This work contributes to this research field by comparing the different algorithms presently deployed and assessing their advantages and shortcomings compared to each other. To this aim, three different thermal-hydraulics problems are considered: two benchmark ’flow over cylinder’ test cases at diverse Reynolds numbers, and the DYNASTY experimental facility operating at Politecnico di Milano, which studies the natural circulation established by internally heated fluids for Generation IV nuclear applications, simulated using the RELAP5 nodal solver. As a key result, this paper highlights the main advantages and disadvantages of the available parametric Dynamic Mode Decomposition methods, concluding that the choice of the algorithm version strongly depends on the problem under consideration and on the user priority (i.e., accuracy or computational speed of the online phase); additionally, this paper shows that an interpretable linear model can be learnt from parametric datasets governed by non-linear models, that can be used for parameter interpolation and, most importantly, for state prediction in time. K eywords DMD \cdot RELAP5 \cdot Dimensionality Reduction \cdot Parametric Model Discovery 1 Introduction In the broad world of computational sciences for engineering applications, there is a need for accurate, reliable and efficient mathematical models for multi-query or real-time scenarios [ 1 , 2 ] , both for design and monitoring purposes [ 3 ] . Nowadays, mathematical modelling is (mainly) carried out through Partial Differential Equations (PDEs), and their discretized version (retrieved adopting the most suitable technique for the application of interest) represents the Full Order Model (FOM), which typically describes, typically in detail, the underlying physics of the investigated system. The numerical solution of this FOM is, however, often computationally expensive, which makes it not suited for all these scenarios requiring multiple in-series solutions (design, parameter optimization, sensitivity analyses) or real-time evaluations of the state of the system (online monitoring, control) [ 4 ] : to address this issue, the Reduced Order Modelling (ROM) framework [ 2 , 5 , 6 ] has been introduced. These methodologies allow retrieving a surrogate (i.e., low dimensional) mathematical model starting from a high-fidelity and highly complex one, to be used for multi-query applications, where multiple evaluations of the state of the dynamical system under different conditions are required, and online monitoring and control, as they greatly reduce the computational time and resources needed to obtain a solution through simulation (to the point that such models, once deployed, can even run on personal laptops). Indeed, in a nutshell, all ROM techniques aim at discovering a reduced representation of the FOM starting from some training solutions, called snapshots , which accurately describe the physical problem. This reduction becomes functional if the computational time to obtain an approximation of the FOM solution for an unseen parameter (including time) is much shorter compared to that of the FOM, keeping, at the same time, the required accuracy. Various ROM techniques exist: among them, the Dynamic Mode Decomposition (DMD) [ 7 , 8 ] has gained popularity in the engineering community due to 1) its simple implementation and 2) its good performance for different applications, which range from fluid dynamics [ 9 ] to plasma physics [ 10 , 11 ] to reactor dynamics [ 12 , 13 , 14 ] . However, the biggest selling point of DMD is that it identifies the dominant structures in the original problem, with which it generates a surrogate linear model based on the starting dataset only, without requiring any knowledge of the governing equations: by retrieving a surrogate linear operator that describes the dataset, it can be used also to advance the dynamical system in time beyond the span of the starting dataset. In fact, DMD defines dimensionality reduction in a pure data-driven (non-intrusive) way, by learning the best linear model underlying the physical phenomena described by the available spatio-temporal data. As said, DMD methods yield to interpretable dynamics which many machine learning models, typically black-boxes, lack [ 15 , 16 , 17 ] ; furthermore, the amount of data required for training is much less with respect to other surrogate modelling methods for time-series data, like DeepONet or foundation models [ 18 , 19 ] . In the nuclear engineering community, it is important to understand how a certain prediction from the mathematical model has been obtained, which is why the choice of DMD over other machine learning methods. However, in its standard formulation, DMD cannot deal with parametric time series, meaning that a different linear model has to be derived for each parameter realisation. Indeed, many systems depend not only on time but also on some parameters, such as the inlet fluid velocity in fluid dynamics, the diffusion coefficient in a nuclear reactor or the ambient temperature in a natural circulation loop. This parametric dependence is relevant in the design phase, where the effect of different parameter values on the operating conditions of the system should be studied, but also during online monitoring and control, as changes from the nominal values may bring the system towards accidental conditions. Thus, both time dynamics and parametric dependence should be learnt to have an ’as-general-as-possible’ surrogate model that can describe the system even for unseen scenarios [ 12 ] . Parametric ROM techniques are currently under study and development, including parametrised versions of DMD. To the best of the authors’ knowledge, the most recent and more in-depth research on this topic can be found in [ 11 , 20 , 21 ] . Each one of these research works presents different versions of a Parametric Dynamic Mode Decomposition (pDMD) algorithm, applied to several parametric dynamical systems. Andreuzzi et al. [ 20 ] first performs dimensionality reduction of the parametric dataset; then, they feed the reduced dataset for each parameter realization to the DMD algorithm, forecasting the reduced snapshots in the future and then reconstructing the parametric dynamics during the online phase using an approach similar to POD with Interpolation [ 22 ] is used to estimate the parametric dynamics during the online phase. Huhn et al. [ 21 ] , instead, collect the eigenvalues and the eigenmodes computed by DMD for each parametric dataset, and interpolate them for the unseen parameters; Faraji

Read the complete text