A non-intrusive data-driven reduced order model for parametrized CFD-DEM numerical simulations

Arash Hajisharifi, Francesco Romano, Michele Girfoglio, Andrea Beccari, Domenico Bonanni, Gianluigi Rozza

TL;DR

This paper introduces a novel non-intrusive data-driven Reduced Order Model (ROM) for computationally expensive CFD-DEM simulations of fluid-solid systems. By combining Proper Orthogonal Decomposition with Interpolation (PODI) and the Finite Volume method, they achieve significant speedups while maintaining good accuracy, paving the way for more efficient design and analysis of industrial processes.

What Problem Does This Paper Address?

Simulating fluid-solid interactions, particularly in industrial processes like fluidized beds, is crucial for design and optimization. However, these simulations, often employing Computational Fluid Dynamics (CFD) coupled with the Discrete Element Method (DEM), are computationally intensive. This high cost severely limits the ability to explore a wide range of design parameters or operating conditions. Existing methods, while improving, still fall short of the extensive simulation needs of industry. The paper aims to bridge this gap by developing a faster, yet accurate, surrogate model that can predict the behavior of these complex systems across various configurations.

Methods

The core of the proposed solution is a non-intrusive data-driven Reduced Order Model (ROM). This means the ROM is built using only the output data from the high-fidelity CFD-DEM simulations (the "Full Order Model" or FOM), without needing access to the FOM's source code.

Key methodological contributions include:

  • PODI (Proper Orthogonal Decomposition with Interpolation): This technique is used to construct a reduced basis for the system's state. POD identifies the dominant modes of variation in the simulation data, and interpolation (specifically Radial Basis Functions - RBFs) is used to map these modes to new parameter values.
  • Finite Volume (FV) Method: Unlike previous works that used Finite Difference (FD) methods, this paper adopts the FV method for spatial discretization. FV is more versatile for complex geometries and is widely used in open-source CFD codes like OpenFOAM, making the ROM approach more practical for industrial applications.
  • Global vs. Local PODI: The paper explores two variants of PODI:
  • Global PODI: A single POD basis is computed from all training snapshots across all parameter values.
  • Local PODI: A separate POD basis is computed for each parameter value in the training set, and RBFs are used to interpolate these bases for new parameter values. This approach is hypothesized to better capture local parameter-dependent behaviors.
  • Parametric Study: The ROM is specifically designed to handle parameter variations. In this work, the Stokes number (Stk), a key dimensionless parameter characterizing particle behavior in a fluid, is used as the primary parameter.

Datasets

The study uses a fluidized bed benchmark problem as its primary dataset. The high-fidelity CFD-DEM simulations provide the "snapshots" (solutions at specific time points) used to train and validate the ROM.

  • Training Data: A significant portion (90%) of the collected FOM snapshots are used for training the ROM.
  • Validation Data: The remaining 10% of FOM snapshots are used to evaluate the ROM's accuracy.
  • Benchmark: The fluidized bed system is a well-established benchmark for CFD-DEM validation.

Key Insights and Ideas

  • FV + ROM Synergy: Combining the robust Finite Volume method with data-driven ROMs is a promising avenue for practical industrial applications, especially for complex geometries where FD methods can be restrictive.
  • Local PODI Superiority: The local PODI approach demonstrates significantly better accuracy and efficiency compared to global PODI, particularly in capturing the dynamics of the fluid phase across different Stokes numbers. This highlights the importance of parameter-specific basis functions.
  • Lagrangian vs. Eulerian Dynamics: The number of POD modes required to capture a certain energy threshold differs significantly between the Eulerian (fluid volume fraction) and Lagrangian (particle position and velocity) fields. Lagrangian velocity, in particular, requires a much larger number of modes, indicating its higher complexity.
  • Stokes Number Significance: The Stokes number is a critical parameter that strongly influences particle behavior. Developing parametric ROMs that can accurately predict system response across a range of Stk values is essential for industrial relevance.

Key Results

  • Time Reconstruction: The ROM successfully reconstructs the time evolution of both Eulerian (fluid volume fraction) and Lagrangian (particle position and velocity) fields with good accuracy. For the Eulerian field, using 90% of the training data and retaining 90% of the energy (150 POD modes) resulted in a mean L2-norm error below 10%. For the Lagrangian field, achieving similar accuracy required a significantly higher number of POD modes.
  • Parametric Study (Stokes Number):
    • The local PODI approach significantly outperformed the global PODI in reconstructing the fluid volume fraction for different Stokes numbers. Global PODI showed spurious oscillations, while local PODI captured main patterns with higher accuracy.
  • The local PODI was also substantially faster, achieving a speedup of about an order of magnitude compared to global PODI.
  • Computational Speedup: The ROM achieved a remarkable global speedup of approximately 2e5 compared to the FOM for simulations. The online computation of reduced coefficients was orders of magnitude faster than the full FOM simulation.

Limitations and Future Work

  • Lagrangian Parametrization: While the paper presents a parametric study for the Eulerian phase with respect to the Stokes number, the parametrization of the Lagrangian phase is left for future work.
  • Non-linear Behavior: The current approach relies on POD, which is primarily linear. The authors acknowledge that capturing complex non-linear behaviors in CFD-DEM might require more advanced techniques.
  • Future Directions:
    • Machine Learning Integration: The authors plan to explore Machine Learning (ML) techniques, such as Autoencoders and Feed Forward Neural Networks, to better capture non-linear dynamics and improve the mapping of reduced coefficients to parameters.
  • Full Parametric ROM: Extending the parametric ROM to include the Lagrangian phase and other relevant parameters is a key future goal.
  • Efficiency and Accuracy Improvement: Continued efforts will focus on enhancing both the efficiency and accuracy of the ROM reconstruction, particularly in parametric frameworks.

Summary updated

Abstract

The investigation of fluid-solid systems is very important in a lot of industrial processes. From a computational point of view, the simulation of such systems is very expensive, especially when a huge number of parametric configurations needs to be studied. In this context, we develop a non-intrusive data-driven reduced order model (ROM) built using the proper orthogonal decomposition with interpolation (PODI) method for Computational Fluid Dynamics (CFD) -- Discrete Element Method (DEM) simulations. The main novelties of the proposed approach rely in (i) the combination of ROM and FV methods, (ii) a numerical sensitivity analysis of the ROM accuracy with respect to the number of POD modes and to the cardinality of the training set and (iii) a parametric study with respect to the Stokes number. We test our ROM on the fluidized bed benchmark problem. The accuracy of the ROM is assessed against results obtained with the FOM both for Eulerian (the fluid volume fraction) and Lagrangian (position and velocity of the particles) quantities. We also discuss the efficiency of our ROM approach.
BibTeX
@article{hajisharifi2023non,
  author = {Arash Hajisharifi and Francesco Romano and Michele Girfoglio and Andrea Beccari and Domenico Bonanni and Gianluigi Rozza},
  title = {{A non-intrusive data-driven reduced order model for parametrized CFD-DEM numerical simulations}},
  journal = {Journal of Computational Physics},
  year = {2023},
  doi = {10.1016/j.jcp.2023.112355},
}
Figure