Technical Summary: OnlyDense – Reduced-Order Modeling for Lagrangian Simulation

Problem Statement

Lagrangian simulation methods, such as Smooth Particle Hydrodynamics (SPH) and the Material Point Method (MPM), are essential for studying dynamic systems in science and engineering, particularly in scenarios involving large deformations, fragmentation, and hypervelocity impacts (e.g., spacecraft shielding against orbital debris). However, these methods face two critical limitations:

1. Computational Cost: The cost of simulation scales quadratically with the number of particles due to the reliance on neighborhood estimation for spatial gradients. Simulations involving millions of particles become prohibitively expensive.
2. Non-differentiability: The discrete, non-differentiable nature of numerical simulations hinders inverse design problems, where finding an optimal configuration requires a full simulation run at every iteration.

While machine learning has been applied to dynamical systems, existing literature primarily focuses on Eulerian (grid-based) simulations or graph-based methods for Lagrangian systems. Graph-based approaches treat the state as a discrete set of particles, tying the complexity of both representation and dynamics directly to the discretization size, which prevents scalable Reduced-Order Modeling (ROM) for massive particle systems.

Methodology

The authors propose OnlyDense, a learning framework that treats the system state as a continuous function rather than a discrete set of particles. The system's trajectory is modeled as a sequence of functions evolving within a Hilbert space.

Core Formulation

Instead of embedding the state in a nonlinear latent manifold or using graph neural networks, the method approximates the state space using a learned linear subspace spanned by a finite set of neural basis functions.

• State Representation: The state $s_t$ at time $t$ is defined over a reference configuration $\Omega_0$ and approximated as:
  $$s_t(X) = \sum_{k=1}^{K} c_k(t) \phi_k(X)$$
  where $\{\phi_k\}_{k=1}^K$ are learned basis functions (parameterized by neural networks) and $c_k(t)$ are latent coefficients.
• Basis Learning (Function Encoder): The authors employ a "Function Encoder" to learn the set of basis functions $\{\phi_k\}$. The coefficients $c(t)$ are obtained via direct projection (inner product) or least-squares optimization, avoiding the need for complex encoders or optimization-based auto-decoding schemes used in prior Implicit Neural Representation (INR) works.
• Dynamics Learning: Once the state is reduced to a low-dimensional coefficient vector $c(t)$, the system dynamics are learned entirely within this reduced space. The framework supports both:
  • Discrete-time: Learning a map $c_{t+1} = f(c_t; \Psi)$.
  • Continuous-time: Learning the time derivative $\dot{c}(t) = f(c, t; \Psi)$ using Neural Ordinary Differential Equations (Neural ODEs).
• Shared Basis: The framework enforces that both position and velocity fields share the same set of basis functions, ensuring physical consistency.

Key Technical Distinctions

• Linear vs. Nonlinear: Unlike recent works that parameterize state functions as nonlinear manifolds, OnlyDense restricts the space to a linear subspace. This allows for explicit access to basis functions and coefficients, enabling direct projection rather than iterative optimization for encoding.
• Scalability: The complexity of the representation is controlled by the number of basis functions $K$, independent of the number of discretization points $N$. The method allows sampling a subset of particles ($M \ll N$) to compute coefficients, making it feasible for simulations with over one million particles.
• Non-Intrusive: The method does not require access to the governing equations, classifying it as a non-intrusive ROM.

Experimental Results

The method was evaluated on both Eulerian benchmarks (2D Wave and Navier-Stokes equations) and large-scale Lagrangian simulations generated using the ABSTRAO solver.

Datasets

1. Eulerian Benchmarks: 2D Wave equation and 2D Navier-Stokes on a $64 \times 64$ grid.
2. Lagrangian Simulations:
   • Exploding Plate: ~22,000 particles.
   • Mott Ring: ~3,400 particles.
   • Projectile Impact: ~1.1 million particles (hypervelocity impact of an aluminum projectile on a steel plate).

Performance Metrics

• Reconstruction Accuracy: The method achieved $R^2$ scores exceeding 0.99 for most datasets using as few as 32 basis functions. For the 1.1 million particle "Projectile" simulation, the $R^2$ score reached approximately 99% with 32 bases.
• Sampling Efficiency: On the 2D Navier-Stokes and Wave equations, OnlyDense demonstrated stable performance across varying spatial sampling rates (5%, 25%, 50%). While some baselines performed well at high sampling rates, their performance degraded significantly under sparse observations, whereas OnlyDense remained robust.
• Comparison: In comparisons with baselines like I-MP-PDE, DeepONet, SIREN, and DINo, OnlyDense provided competitive or superior accuracy, particularly in out-of-time (temporal extrapolation) scenarios and under sparse sampling conditions.

Significance and Claims

The paper positions OnlyDense as a unification of classical projection-based Reduced-Order Modeling (specifically Proper Orthogonal Decomposition, or POD) with modern deep learning.

• Conceptual Bridge: The approach offers a conceptual bridge between POD and deep learning. While POD relies on Principal Component Analysis (PCA) which scales poorly with resolution and is limited to Eulerian settings, OnlyDense replaces discretized spatial modes with learned neural basis functions. This allows evaluation at arbitrary locations and extends ROM to Lagrangian systems with irregular geometries.
• Interpretability: The latent variables correspond directly to coefficients in Hilbert space, and the basis functions correspond to spatial modes, providing a natural interpretation analogous to classical modal analysis.
• Scalability: By decoupling the representation complexity from the discretization size, the method enables accurate reconstruction and dynamics prediction for massive particle systems (millions of particles) that are computationally intractable for traditional graph-based ROM approaches.

Limitations

The authors explicitly acknowledge a current limitation: the method relies on a single reference configuration. While effective for Eulerian simulations (fixed domain) or Lagrangian simulations with a single reference, the framework cannot yet generalize to scenarios involving multiple simulations with different reference configurations or highly irregular geometries that change topology significantly. This is identified as a primary target for future work.