Model Reduction for the Material Point Method on Nonlinear Manifolds Using Deep Learning

Peter Yichen Chen, Maurizio Chiaramonte, Eitan Grinspun, Kevin CarlbergAAAI 2021 · OpenReview

Introduction

The Material Point Method (MPM) is a powerful hybrid Lagrangian-Eulerian technique widely used in computer graphics and engineering to simulate complex materials like snow, foam, and fluids. While MPM excels at handling large deformations and topological changes, its computational cost is prohibitive for real-time applications or rapid design iterations. This paper introduces a novel Reduced-Order Model (ROM) that captures the essential physics of MPM simulations on a low-dimensional, nonlinear manifold learned via deep learning, achieving significant speedups with minimal loss in accuracy.

Methodology: Learning the Kinematic Manifold

Instead of tracking thousands of individual material points, the proposed method defines the "shape" of possible deformations using a neural network. A 5-layer Multi-Layer Perceptron (MLP) acts as a decoder, mapping a small set of generalized coordinates to the full-scale deformation of the object.

The mapping function f relates the initial reference state to the deformed state on the learned manifold.

Gradient-Aware Training

A critical innovation in this work is the use of a custom loss function during training. To ensure physical consistency, the authors do not merely match point positions; they incorporate penalty terms ((\lambda_1, \lambda_2)) for the deformation gradient and velocity. This "gradient-aware" approach prevents the model from accumulating massive errors during dynamic motion.

Hyper-reduction and Efficiency

To achieve true computational savings, the authors implement "hyper-reduction" techniques. Specifically, the "Sample-node centric" approach utilizes quadrature sampling on the background grid. By inverting the neural network mapping, the system can identify where these points originated in the reference material, making the simulation cost independent of the total particle count.

Key Insights: The Importance of Gradients

The researchers found that simply matching positions in training snapshots was insufficient for stable simulations. As shown in the training parameter analysis, incorporating velocity and deformation gradient penalties is essential for reducing error across all metrics.

Table showing how non-zero penalty terms for velocity and deformation gradients significantly reduce simulation errors.

By ensuring the manifold mapping is injective (one-to-one), the model can accurately "look up" reference positions from deformed coordinates. This allows for the generation of effective material points on the fly, bypassing the need to track the original high-resolution particle set.

Results and Performance

The method was evaluated using a "poking test" on an elastic cylinder (1,368 material points, (E=12500) Pa, (\nu=0.3)). Key results include:

  • Dimensionality Reduction: The simulation was compressed from 1,368 material points to just 5 generalized coordinates—a reduction factor of 821x.
  • High Fidelity: The model maintained a position error of only 0.55% on unseen testing data.
  • Computational Speedup: Using hyper-reduction, the dynamics were computed using only 15 material points, demonstrating the potential for orders-of-magnitude faster execution compared to full MPM.

Conclusion and Future Work

This research demonstrates that deep learning can effectively learn the nonlinear manifolds inherent in complex material physics. While currently focused on elastic materials, future work aims to extend this framework to handle plasticity, fracture, and complex topological changes. The ability to represent high-fidelity simulations with just a handful of coordinates opens new doors for interactive physics and real-time engineering tools.

Summary updated

BibTeX
@inproceedings{chen2021model,
  author = {Peter Yichen Chen and Maurizio Chiaramonte and Eitan Grinspun and Kevin Carlberg},
  title = {{Model Reduction for the Material Point Method on Nonlinear Manifolds Using Deep Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=LuKzNwz_0y},
}
Figure