Comparison of Data-Driven Approaches to Modeling Complex Behavior of 2D Liquid Simulator

Dmitry Kovalev, Dmitry Khliustov, Sergey O. SafonovDAMDID/RCDL (Supplementary Proceedings) 2021

Introduction

Modeling the dynamics of complex systems, such as 2D liquids, has traditionally relied on Partial Differential Equations (PDEs). While these hand-crafted models are physically accurate, they are computationally expensive and require significant expertise to derive. The rise of data-driven modeling offers a promising alternative: learning fluid dynamics directly from observational data. However, liquid simulation presents a significant challenge due to its high dimensionality, non-linear interactions, and the need for models that generalize across different initial conditions. This paper evaluates several neural architectures to determine which best captures these intricate behaviors.

Methodology

The researchers compared five distinct architectures, categorized primarily by their "locality"—whether they model global system states or local particle interactions.

Comparison of data-driven modeling approaches.
  1. Dynamic Mode Decomposition (DMD): A linear algebra technique that identifies a linear operator to describe the system's evolution over time.
  2. Graph Neural Networks (GNN): A local approach treating particles as nodes. It assumes a particle's next state depends only on its $k$-nearest neighbors, mimicking physical locality.
  3. Neural Ordinary Differential Equations (NODE): A continuous-time model that learns the derivative of the system's state.
  4. NODE with DMD (NODEwD): A hybrid model where DMD decomposes high-dimensional data into spectral coefficients, which the NODE then predicts.
  5. CNN with DMD (CNNwD): Similar to NODEwD, but utilizes a Convolutional Neural Network to predict DMD coefficients.

Experimental Setup

The models were tested using a custom dataset generated by a 2D liquid simulator based on smoothed particle hydrodynamics. The dataset included simulations of up to 1,500 particles across 100 different configurations with randomized emitter coordinates and velocities. Models were trained on a single simulation and evaluated on the remaining 99 to test generalizability.

Key Insights: Accuracy vs. Generalizability

A central finding of the study is the trade-off between reconstruction accuracy and the ability to generalize to new scenarios.

  • Decomposition and Denoising: Hybrid models like NODEwD use DMD as a "denoiser," focusing on global patterns. This leads to the lowest error rates and visually smoother trajectories. However, these models are often "brittle," learning specific training scene patterns rather than universal physics.
  • The Power of Locality: GNNs proved to be the most robust. By learning local particle-to-particle interactions, they generalized far better when emitters were moved to new, unseen locations.
  • Metric Selection: The authors highlight that standard Mean Squared Error (MSE) is often insufficient for fluid dynamics. They propose using distribution-based metrics like Optimal Transport (OT) and Maximum Mean Discrepancy (MMD) to better capture visual fidelity.

Results

  • Performance: NODEwD achieved the highest accuracy with the lowest MSE (0.19 ± 0.10).
  • Efficiency: DMD was the fastest method. Notably, adding DMD decomposition to NODE or CNN architectures reduced training and inference times by approximately 5x compared to their standalone versions.
  • Visual Fidelity: While pure DMD struggled with complex patterns, the hybrid NODEwD model reconstructed fluid behaviors that were visually indistinguishable from the ground truth.

Limitations and Future Work

A significant hurdle remains the "rollout problem." While models perform well on one-step predictions, errors compound exponentially when the model uses its own previous predictions to forecast the next step (rollout mode).

Comparison of one-step vs. rollout prediction accuracy.

Future research will focus on improving long-term stability and handling the stochasticity inherent in real-world sensor data. The authors suggest that the most effective future models may be "aggregated" architectures that combine the global pattern-matching strengths of DMD with the local physical reasoning of GNNs.

Summary updated

BibTeX
@inproceedings{kovalev2021comparison,
  author = {Dmitry Kovalev and Dmitry Khliustov and Sergey O. Safonov},
  title = {{Comparison of Data-Driven Approaches to Modeling Complex Behavior of 2D Liquid Simulator}},
  booktitle = {DAMDID/RCDL (Supplementary Proceedings)},
  year = {2021},
}
Figure