Simulator Pipeline¶

Our MPM simulator adopts the explicit time-integration scheme.

The conventional MPM pipeline looks like:

Particles-to-Grid (P2G). Transfer mass and momentum from particles to grid nodes: $\{m_p, m_p v_p^n\} \rightarrow \{m_i, m_i v_i^n\}$
Grid Update. Update grid velocities with either explicit or implicit time integration: $v_i^{n} \rightarrow v_i^{n+1}$
Grid-to-Particles (G2P) and Particle Advection. Transfer velocities from grid nodes to particles, evolve particle strains, and project particle deformation gradients for plasticity (if any). Update the particle positions with their new velocities: $\{v_i^{n+1}\} \rightarrow \{v_p^{n+1}, F_p^{n+1}\}$ , $\{p_p^{n}, v_p^{n+1}\} \rightarrow \{p_p^{n+1}\}$
Partition Update. Maintain the sparse data structure topology by updating the active-block array and the mapping from block coordinates to array indices.

We propose a novel pipeline with a fused G2P2G kernel as a substitute for the original P2G and G2P. It allows less memory footprint for each particle (although in total doubled due to double-buffering required by G2P2G) and a better performance in general.