A Comparative Study of Parametric Dynamic Mode Decomposition Algorithms on Thermal-Hydraulics Case Studies

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

This paper, recently published by researchers from Politecnico di Milano and UNIST, provides a rigorous "head-to-head" comparison of different Parametric Dynamic Mode Decomposition (pDMD) algorithms. It specifically evaluates how well these data-driven methods can predict complex fluid behavior in nuclear and thermal-hydraulic systems without needing to run expensive, full-scale simulations every time a parameter changes.

TL;DR

The researchers compared four flavors of pDMD to determine which is best for predicting fluid dynamics across varying parameters (like flow speed or heating power). They found that while "Monolithic" approaches are often the most accurate, "Operator Interpolation" methods are significantly faster for real-time applications, with the choice ultimately depending on whether the user prioritizes precision or computational speed.

What Problem Does This Paper Address?

In engineering, we often use Reduced Order Models (ROMs) to create fast "surrogates" for slow, complex simulations (Full Order Models). Dynamic Mode Decomposition (DMD) is a popular tool for this because it’s "equation-free"—it learns the physics directly from data.

However, standard DMD has a major flaw: it’s "one-off." If you train it on a fluid flowing at 1 m/s, it can't accurately predict what happens at 2 m/s. You would traditionally need to build a brand-new model for every single scenario. Parametric DMD (pDMD) aims to solve this by learning how the system's dynamics change as a function of external parameters, allowing it to interpolate and predict behavior for conditions it has never seen before.

Methods

The authors compared three main architectural approaches to pDMD:

  1. Reduced Operators Interpolation (ROI): Proposed by the authors, this method builds DMD models for several training points, "unfolds" the resulting state matrices into a new dataset, and uses a second Singular Value Decomposition (SVD) to find the underlying parametric patterns.
  2. Reduced Koopman Operator Interpolation (RKOI): This method interpolates the actual "modes" (shapes), frequencies, and amplitudes of the system. It utilizes Optimized DMD, which is more robust against noise than the standard version.
  3. Latent Space Interpolation (via the pyDMD library):
    • Monolithic: Stacks all data from all parameters into one giant matrix to find a single, universal operator.
  4. Partitioned: Builds separate models for each parameter and then interpolates the results in the "latent" (compressed) space.

Datasets

The algorithms were tested against three distinct thermal-hydraulic challenges:

  • Laminar Flow Over a Cylinder (FEniCS): A classic benchmark using the dolfinx solver, focusing on simple vortex shedding at Reynolds numbers between 100 and 150.
  • CFDbench Flow Over Cylinder: A more complex version of the above with higher Reynolds numbers (600–1000), introducing more non-linear, oscillatory dynamics.
  • DYNASTY Experimental Facility: Data from a real-world natural circulation loop at Politecnico di Milano, simulated using the RELAP5 nodal solver. This case tested the models' ability to handle both parameter interpolation (heating power) and time-forecasting.

Key Insights and Ideas

  • The Offline/Online Split: ROI and RKOI do the heavy lifting during the "offline" training phase. This makes their "online" (prediction) phase incredibly fast—ideal for real-time monitoring on a laptop or even an edge device.
  • Handling "Chaos": For systems with complex oscillations (like the CFDbench case), RKOI outperformed ROI. This is because RKOI's use of Optimized DMD is better at "catching" the specific frequencies of vortex shedding.
  • The "Rank" Bottleneck: All pDMD methods are highly sensitive to the SVD "rank" (the number of modes kept). If the rank is too low, you miss the physics; if it's too high, the interpolation between parameters becomes noisy and inaccurate.
  • Interpretable AI: Unlike "black-box" neural networks, these pDMD models provide linear operators and modes that engineers can actually inspect to understand the underlying physical frequencies and spatial structures.

Key Results

  • Accuracy: In the DYNASTY case, the Monolithic approach was the most accurate (Relative Error ~0.01), though all four methods provided good reconstructions.
  • Speed: The ROI and RKOI methods were orders of magnitude faster in the online phase. For the DYNASTY facility, ROI took only 0.003 seconds to generate a prediction, compared to 3.2 seconds for the Monolithic approach.
  • Complex Dynamics: In the high-Reynolds number CFDbench test, RKOI achieved an error of 0.06 with a rank of only 10, while ROI struggled (error 0.15) until the rank was significantly increased.

Limitations and Future Work

The authors acknowledge that hyperparameter tuning (specifically choosing the SVD rank) remains a manual, trial-and-error process. A model that works perfectly for one Reynolds number range might fail if the rank isn't re-calibrated for another.

Future Directions: * Data Assimilation: Integrating these pDMD models into a Kalman Filter framework to correct predictions in real-time using actual sensor data. * Multi-Physics: Extending the comparison to coupled problems, such as systems where nuclear reactor power (neutronics) and fluid flow (hydraulics) influence each other simultaneously.

Summary updated

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 state operator of the underlying dynamical system that can be used, in principle, to advance the original dataset in time even beyond its span. 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 parameter realization. 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, whose datasets are, respectively, obtained with the FEniCS finite element solver and retrieved from the CFDbench dataset, 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, whose dataset was generated using the RELAP5 nodal solver.
BibTeX
@inproceedings{riva2026comparative,
  author = {Stefano Riva and Andrea Missaglia and Carolina Introini and In Cheol Bang and Antonio Cammi},
  title = {{A Comparative Study of Parametric Dynamic Mode Decomposition Algorithms on Thermal-Hydraulics Case Studies}},
  booktitle = {Nuclear},
  year = {2026},
  doi = {10.1080/00295450.2025.2575701},
}
Figure