12 Mar 2026 · 2 min · warp PR #1271, closed

Contact lists that remember

bonded-particle DEM needs pair state that survives the substep. I proposed it upstream; it wasn't taken. The idea still matters.

I proposed a persistent contact list for Warp: a data structure for bonded-particle DEM where contacts are built once and updated incrementally, with per-contact state travelling along. The PR was not taken upstream, and that is fine; this post is about why the idea matters anyway, and what I would do differently.

The problem

The standard particle-simulation loop rebuilds its neighbor set every substep from a hash grid. For plain repulsive contact that is correct and simple: pairs need no memory, so recomputing them loses nothing.

Bonded materials break this. A bond has a rest length set at formation. Strain is displacement relative to that remembered length, and a bond that exceeds its strain threshold breaks permanently. Rebuild the pair set from scratch and every pair arrives with no memory: it cannot know its rest length, so it cannot know it is stretched, so it can never break. Fracture is history, and a memoryless loop cannot represent history.

contacts = ContactList(hash_grid, radius)
contacts.rest_length = dist(contacts.pairs)     # set at bond time
contacts.bonded = ones(len(contacts))
for step in range(substeps):
    contacts.update()                           # drop separated, add new, CARRY state
    strain = (dist(contacts.pairs) - contacts.rest_length) / contacts.rest_length
    contacts.bonded &= strain < strain_max      # once broken, permanently broken

The same shape appears anywhere contact has memory: stick-slip friction that must remember where sticking began, adhesion that holds until it does not, plastic deformation, which is path dependence by definition.

Why the GPU makes it interesting

The pair set changes every step, so the layout must support insertion and removal without reallocation, and the per-contact payload has to be permuted alongside the pairs. I used a flat CSR-like layout so pairs and their floats move together. The validation that matters is a reversibility test: load below threshold and release, the material returns to rest; load past threshold, and it must never heal. A bond model that heals is the failure mode to check first.

What I would do differently

The PR was 967 lines proposing a new core data structure, when the maintainer-suggested route was a pure-Warp example demonstrating the pattern. Ship the example first, let the abstraction earn its way in later. Big proposal, small appetite: my mistake, and a reusable lesson about contributing to mature codebases.