Revealing Low-Dimensional Structure in 2D Richtmyer-Meshkov Instabilities via Parametric Reduced-Order Modeling

Daniel A. Messenger, Daniel A. Serino, Balasubramanya Nadiga, Marc KlaskySSRN 2025 · DOI

TL;DR

This paper introduces a novel data-driven framework for modeling the complex Richtmyer-Meshkov Instability (RMI) in 2D. By using a shallow autoencoder and a linear dynamical system in a low-dimensional latent space, they demonstrate that RMI, even in its nonlinear regime, can be surprisingly well-represented by a low-dimensional linear system, offering a significant leap in computational efficiency for simulations.

What Problem Does This Paper Address?

The Richtmyer-Meshkov Instability (RMI) is a crucial phenomenon in various engineering applications, particularly in Inertial Confinement Fusion (ICF). RMI can lead to mixing of materials, degrading the efficiency of fusion reactions. Accurately simulating RMI, especially in its nonlinear growth phase and across a wide range of parameters (like material properties and initial conditions), is computationally expensive and challenging. Existing methods often rely on full-order simulations or simplified models that don't capture the full complexity or are not easily parametrized. This paper aims to bridge this gap by developing an efficient, data-driven Reduced-Order Model (ROM) that can accurately predict RMI evolution across a high-dimensional parameter space.

Methods

The core of the proposed method is a Parametrized Reduced-Order Model (ROM) built upon the Latent Space Dynamics Identification (LaSDI) framework. Key components include:

  • Autoencoder (AE): A neural network used for compression and decompression. It maps high-dimensional input data (material interface bitmaps) to a low-dimensional latent space and reconstructs it. This nonlinear mapping is crucial for capturing the complex interface dynamics.
  • Latent Space Dynamics: A linear dynamical system is imposed on the latent space representation. This is a significant finding, as it suggests that despite the complex physical process, the underlying dynamics in the latent space can be modeled linearly.
  • Parameters-to-Coefficients Map (M): Another neural network that maps the physical parameters (EOS parameters, initial conditions) directly to the coefficients of the latent space dynamics and the initial latent state. This allows for interpolation and prediction for new parameter values without re-running full simulations.
  • POD Initialization: The weights of the autoencoder are initialized using Proper Orthogonal Decomposition (POD) modes of the training data. This significantly improves training efficiency and accuracy compared to random initialization.

Datasets

The study uses synthetic data generated from full-order hydrodynamic simulations of RMI. The simulations cover:

  • A weak coupling regime with 4 dominant wavemodes.
  • A strong coupling regime with 2 dominant wavemodes.
  • A mixed regime combining both.

In total, 1536 simulations were performed across a nine-dimensional parameter space. The data was split into training, validation, and testing sets (approximately 0.725-0.025-0.25 proportions).

Key Insights and Ideas

  • Surprising Linearity in Latent Space: The most striking insight is that the complex, nonlinear evolution of the RMI interface, even into the nonlinear growth regime, can be effectively captured by a linear dynamical system in a low-dimensional latent space. This suggests a fundamental low-dimensional structure underlying the RMI.
  • Autoencoder as a Key to Dimensionality Reduction: The nonlinear mapping provided by the autoencoder is essential for compressing the high-dimensional interface data into a manageable latent space where linear dynamics can then be applied.
  • Parametric Interpolation is Feasible with Neural Networks: Traditional interpolation methods struggled with the high-dimensional parameter space. The paper demonstrates that a neural network-based parameters-to-coefficients map can successfully interpolate and predict RMI behavior across this space.
  • POD Initialization Boosts Performance: Initializing the autoencoder weights with POD modes significantly accelerates training and improves the accuracy and generalization of the ROM.

Key Results

  • Accurate Prediction: The proposed F-ROM (Full Reduced-Order Model) achieves high accuracy in predicting RMI evolution. For the weak coupling regime, the 95th percentile Mean Pixel Error (MPE) on the test set was as low as 3.7% and Jaccard Loss (JL) was 9.9% with a 3-dimensional latent space ($n_z=3$).
  • Optimal Latent Dimension: The study consistently found that a 3-dimensional latent space ($n_z=3$) offered the best balance between accuracy and generalization across different coupling regimes. Higher dimensions ($n_z=4$) often led to overfitting and poorer generalization in the F-ROM.
  • Efficiency: The ROM framework allows for rapid simulation of RMI dynamics across a wide parameter range, significantly reducing computational cost compared to full-order simulations.
  • Generalization: The F-ROM demonstrated good generalization capabilities to unseen parameter values and initial conditions.

Limitations and Future Work

  • Model Architecture: While a shallow autoencoder proved effective, the authors suggest that deeper architectures or more complex latent space dynamics could potentially improve performance, especially for modeling the strongly nonlinear growth phase.
  • Parameter Space Coverage: The current study focused on specific parameter ranges. Future work could explore a wider range of EOS parameters and initial conditions.
  • Incorporating Dominant Modes: Visualizations of the latent space suggest that simulations cluster based on dominant interface modes. Future work aims to directly incorporate this information into the parameters-to-coefficients map.
  • Theoretical Foundation: The authors aim to provide theoretical evidence for the observed drastic dimensionality reduction in future work.
  • Combining Techniques: The paper suggests combining their framework with techniques for inferring physical unknowns from radiographic images to improve inference and control capabilities in experiments.

Summary updated

BibTeX
@misc{messenger2025revealing,
  author = {Daniel A. Messenger and Daniel A. Serino and Balasubramanya Nadiga and Marc Klasky},
  title = {{Revealing Low-Dimensional Structure in 2D Richtmyer-Meshkov Instabilities via Parametric Reduced-Order Modeling}},
  year = {2025},
  doi = {10.2139/ssrn.5741330},
  howpublished = {SSRN},
}
Figure