Large-N instability fixed + validation suite as UI presets + it runs in a browser

[algorithm0r]

Hi Anthony — Chris here (with an AI assistant doing the legwork). You saw the demo videos; this is the code behind them — and as of tonight you can run the whole thing in a browser:

▶ https://algorithm0r.github.io/Physics-Sim/

Chrome/Edge 113+ with hardware acceleration. Click the canvas once, then Space. The Presets menu holds the entire validation suite (T0–T10) plus two demo scenes — each loads the exact configuration the tests validated.

Your simulator was too good to leave stalled, so we picked up where you left off and went after the large-N blow-up. Nothing here is a rewrite — it's your architecture throughout, with surgical fixes.

The instability (what it actually was)

H3 — grid cell capacity collapse (the big one). grid_capacity() sizes cells from min_rad, but with variable_rad=false min_rad collapses to max_radius, so polydisperse scenes got cell_cap ≈ 11 instead of ~51. Overfull cells silently dropped grid insertions → missed contacts → energy injection that grew with N. Fix: derive variable_rad from the actual radius spread. Cell overflow at 10k particles went 3758 → 0 and total energy stopped growing at every N tested (up to 50k).

H8 — broad-phase clear/insert race. Clearing and inserting the grid in one dispatch races across workgroups; binning is now a third dispatch (2D_Grid_Insert.wgsl).

H1 — local damping. Your commented-out Cundall damping re-enabled behind a setting; the commented moment line needed abs(moment) or it injects rotational energy.

D1 — PFC-style viscous contact dashpot (new, gated by dashpot_beta, 0 = exactly legacy, proven byte-identical). Measured restitution matches the no-tension spring-dashpot ODE to RMS 3e-4.

Validation

Quantitative gates against closed-form/independent references: two-body suite (impact, oblique stick/slip kink at the predicted angle, slide→roll, angular momentum with unequal radii), bonded cantilever vs discrete beam theory (<0.7%), 50k-particle settling with strictly monotone energy decay, bulk tests (repose, packing fraction, oedometer modulus). Every stage re-verified by independent re-runs from artifacts.

The web port

Your index.html/index.js groundwork carried it most of the way. The rest was a strictness gauntlet — none of it logic errors, all things native wgpu 0.16 never enforced: @interpolate(flat) on integer vertex outputs; no writable storage in vertex stages; requiredLimits never forwarded by the 0.16 web backend (shimmed in index.js); browsers cap bind groups at 4 and storage buffers at 16/stage, so all pipelines got regrouped (your group 0–2 structure survived — the small uniforms merged into group 3); plus a few Tint pedantries. After every shader-touching change we re-ran a 1.2M-step bonded fracture scenario and required byte-identical output — the physics held through all of it.

Also in here: headless mode (--headless scenario.json out.csv) that drives your pipeline windowless and dumps per-step state — the whole Python validation harness runs against it (separate repo, happy to share).

Genuinely: this codebase was a pleasure to work in. The bones were right; it was one sizing bug away from working at scale.

🤖 Generated with Claude Code

[algorithm0r]

Status note on the wasm commit (0f02c1e): builds and boots, but browser rendering is not yet verified — don't sink time into pkg/ yet. Verified so far: compiles for wasm32, WebGPU adapter acquired, winit event loop running, no panics in console, presets embedded. Not yet verified: an actual rendered frame (headless Chrome can't present WebGPU reliably, so the pixel check needs a real browser — happening shortly). Will follow up here either confirming it renders or with the fix. Native is unaffected: the harness exe was re-verified byte-identical after every change in that commit.

🤖 Generated with Claude Code

[algorithm0r]

Follow-up on the wasm status note: the web build now renders and runs, live at https://algorithm0r.github.io/Physics-Sim/ (Chrome/Edge 113+, hardware acceleration on — WebGPU needs a secure context, which Pages provides).

Getting there took a strictness gauntlet that might interest you, since none of it was wrong logic — just things native wgpu 0.16 never enforced and the browser does: @interpolate(flat) on integer vertex outputs, no writable storage in vertex stages, requiredLimits never forwarded by the 0.16 web backend (shimmed in index.js), the big one — browsers cap bind groups at 4 and storage buffers at 16/stage, so every pipeline's bindings got regrouped (your group 0–2 structure survived; the small uniforms merged into group 3) — and a few Tint pedantries (storageBarrier uniformity, atomicStore, a var length shadowing the builtin).

The physics is provably untouched: after every shader-touching change we re-ran a 1.2M-step bonded fracture scenario and required byte-identical output against the validated baseline — it held through all of it (commits 096cf57, b2701ec).

Try the Presets menu in the browser — the whole validation suite plus the demo scenes, one click each. Click the canvas once, then Space to run.

🤖 Generated with Claude Code