When simulating physical materials that undergo extreme shape changes—such as a splashing liquid, tearing plasticine, or a collapsing snow avalanche—traditional numerical methods face a fundamental dilemma. Computational physicists typically choose between two distinct perspectives:
- The Lagrangian View (Particle-Based): You track discrete points or elements moving through space. This makes it easy to conserve mass and track sharp material boundaries. However, if the material deforms heavily, the underlying mesh connecting these points becomes tangled and distorted, causing the simulation to crash. Purely mesh-free particle methods avoid tangling but require expensive, dynamic neighbor searches to calculate forces.
- The Eulerian View (Grid-Based): You look at fixed coordinates in space and observe how material flows through them, much like watching water pass under a bridge. This is highly efficient for computing spatial derivatives (like pressure gradients), but it struggles to track sharp material boundaries or the historical deformation of a solid.
To overcome these limitations, hybrid Eulerian-Lagrangian frameworks—most notably the Material Point Method (MPM)—combine the best of both worlds. By dynamically transferring information between moving particles and a fixed background grid, these methods handle large deformations, topological fractures, and multi-material interactions with remarkable stability.
The Core Mechanism: The Particle-Grid Loop
The defining feature of a hybrid particle-grid simulation is that it does not choose one perspective; instead, it cycles between them every single time step. The simulation state is permanently stored on Lagrangian particles (or material points), which carry mass, momentum, velocity, and deformation history. The Eulerian grid is temporary, serving as a scratchpad to compute forces and spatial gradients.
A single time step follows a four-stage loop:
[Particles] --- 1. P2G Transfer ---> [Background Grid]
^ |
| 2. Solve Equations
| |
4. Advect Particles <-- 3. G2P Transfer <----+
- Particles-to-Grid (P2G): At the beginning of the step, particle properties (like mass and momentum) are interpolated onto the nodes of a regular background grid using smooth weight functions.
- Solve on the Grid: The governing equations of motion (such as conservation of momentum) are solved on the grid nodes. Because the grid is perfectly structured and uniform, computing spatial gradients—such as the divergence of internal stress or collision forces—is straightforward, stable, and computationally cheap.
- Grid-to-Particles (G2P): The updated velocities, accelerations, and forces calculated on the grid are interpolated back to the particles.
- Advect Particles: The particles update their positions and internal deformation states using the new grid-derived values. Crucially, the background grid is then completely reset or discarded.
Because the grid is reset every step, it never becomes distorted or tangled, no matter how wildly the particles move. As noted in a comprehensive survey by Wang et al., this hybrid coupling makes MPM exceptionally robust for simulating complex interactions between fluids and highly deformable solids (Wang et al., 2024).
Why It Works: Intuition and Advantages
Natural Handling of Fractures and Topological Changes
In traditional mesh-based methods, splitting an object in half requires complex remeshing algorithms to redefine the boundaries. In a hybrid particle-grid framework, separation happens naturally. If two groups of particles move away from each other, they simply stop interpolating to the same grid nodes.
This capability is why the method is widely used to simulate destructive geophysical phenomena. For example, Gaume et al. and Trottet et al. utilized hybrid Eulerian-Lagrangian simulations to model the structural collapse of porous snow layers and the subsequent transition of an avalanche from a solid slab into a fluid-like granular flow (Gaume et al., 2018; Li et al., 2020; Trottet et al., 2022). Similarly, Wolper et al. applied this framework to capture the violent brittle fracture of calving icebergs and their immediate hydrodynamic interaction with the ocean (Wolper et al., 2021).
Bypassing the $O(N^2)$ Neighbor Search
Pure particle methods like Smoothed Particle Hydrodynamics (SPH) require calculating forces between nearby particles. As particles move, finding their neighbors requires building and updating expensive tree structures or spatial hash tables.
By using a fixed grid as an intermediary, hybrid methods eliminate this overhead. Particles only ever interact with the grid nodes immediately surrounding them. As demonstrated by Sharabi et al., shifting the core spatial processing to a regular voxel grid avoids the high computational costs of graph-based particle models, allowing standard image-to-image neural architectures (like U-Nets) to act as highly efficient physics emulators (Sharabi et al., 2024; Sharabi, 2025).
Differentiable Physics and Inverse Problems
Because the mathematical functions used to transfer data between particles and the grid are smooth and continuous, the entire simulation pipeline can be made differentiable. This means you can compute analytical gradients of the simulation's outputs with respect to its inputs.
Differentiable simulation allows researchers to use gradient-based optimization to solve complex "inverse problems"—such as finding the exact material properties of an observed object or designing the control inputs for a robot.
- Soft Robotics Co-Design: Hu et al. developed ChainQueen, a differentiable simulator based on the Moving Least Squares Material Point Method (MLS-MPM). Because the simulator provides exact gradients, optimization algorithms can simultaneously refine a soft robot's physical shape and its walking gait far faster than traditional reinforcement learning algorithms (Hu et al., 2018; Spielberg et al., 2023).
- Deformable Manipulation: The PlasticineLab benchmark by Huang et al. leverages this differentiability to optimize robotic trajectories for manipulating highly plastic materials (Huang et al., 2021). This has been extended to force-aware tasks like robotic food cutting, where Koh et al. tracked impulse exchanges between particles and grids to estimate transient contact forces during material fracture (Koh et al., 2026).
Bridging Vision and Physical Realism
The dual nature of hybrid particle-grid representations aligns perfectly with modern computer vision and neural rendering pipelines. In computer graphics, grids are ideal for volumetric rendering, while particles are ideal for tracking physical dynamics.
By coupling these representations, researchers can infer physical properties directly from video. For example, Li et al. introduced PAC-NeRF, which maps density from an Eulerian voxel grid used for neural radiance fields to Lagrangian particles for physics simulation, ensuring that reconstructed objects obey the conservation laws of continuum mechanics (Li et al., 2023).
Similarly, Lee et al. combined 3D Gaussian Splatting with an MPM simulator to create robust human avatars with realistic clothing physics (Lee et al., 2025), while Lei et al. coupled Lagrangian object particles with Eulerian wind fields to reconstruct invisible environmental forces from sparse video observations (Lei et al., 2026).
Whether applied to tracking moving interfaces (Zhao, 2009), processing unstructured point clouds (He et al., 2020), or modeling geophysical flows (Cotter et al., 2004), the core intuition remains the same: by decoupling the tracking of mass (particles) from the calculation of forces (grids), hybrid simulation bypasses the geometric constraints that cause single-representation solvers to fail.
Key Foundational Work
Hamiltonian Particle-Mesh Method for Two-Layer Shallow-Water Equations Subject to the Rigid-Lid Approximation (2004) — contributed 1 concept elements to this direction.
The remapped particle‐mesh semi‐Lagrangian advection scheme (2007) — contributed 1 concept elements to this direction.