Skip to content

Realistic substrates, solved once

Walk once. Replay forever.

A realistic substrate is a mesh — a .ply from CACTUS, DisCO, or your own builder. A compatible generator walks it once, and we freeze that walk into a replay pack: a lossless intermediate that decodes any acquisition instantly. The payoff — mesh substrates become as fast to evaluate as analytical sphere-and-cylinder models, with none of their acquisition-scheme limits, and the people who built the substrate get their name on a citable DOI.

How it works Bring a machine Read the .rpk spec

① Substrate builders ② Generators ③ The record · DOI ④ Decode anything CACTUS DisCO …any mesher .ply dmipy-sim MC-DC COMING SOON DOI zenodo · minted .rpk PAYLOAD Substrate — single-bundle axons Author · the substrate builder Manifest · shard contributors + everyone who gave walkers PGSE · OGSE · STE real RF · B₁ pulses any b, Δ, δ any B₀ · χ T₂ · T₁ · ρ · MT
Fixed at walk time: geometry · intrinsic diffusivity · permeability. Everything on the right is a replay knob — any waveform, any real B₁ pulse, no short-pulse approximation.

As fast as analytical. As real as a mesh.

Microstructure modelling has always faced a trade-off. Analytical models — signals for spheres and cylinders — are fast, but they idealise the geometry and lean on acquisition approximations (short-pulse, Gaussian-phase) that only hold for some sequences. Monte-Carlo on a realistic mesh is faithful to any sequence, but slow — too slow to sit inside a fit.

The replay pack closes the gap. Pay the Monte-Carlo cost once, freeze the walk, and every later evaluation is a cheap contraction against the stored trajectories:

  • Analytical-model speed

    A decode is a matrix contraction, not a simulation. A solved substrate answers a fit's inner loop as fast as a closed-form sphere/cylinder expression — but the geometry underneath is a real mesh.

  • No acquisition limits

    Because it replays the actual walk, there is no SGP, no Gaussian-phase, no fixed pulse shape. Arbitrary G(t), real B₁ pulses, OGSE, STE — all exact to the Monte-Carlo floor.

  • Builders get the credit

    The substrate author is named on the pack's DOI; everyone who donated walkers is in its manifest. The Commons exists to make building a good substrate a citable contribution.

  • Solved, then permanent

    Anyone can lower a substrate's noise floor from a laptop or a free GPU; shards are additive and never regress. Once it's converged to σ* and walked to 200 ms, it's frozen and DOI'd — solved forever.

The four moves

  • Bring a mesh

    Any open geometry — a .ply from CACTUS, DisCO, or your own builder. The substrate carries its own orientation frame, so gradients and B₀ are defined relative to the fibre.

    The IR

  • Walk it once

    A compatible generator — dmipy-sim today, MC-DC soon — walks the geometry and records trajectories, boundary contact, and the susceptibility field. Intrinsic diffusivity and permeability are fixed here.

    The replay invariant

  • Decode anything

    Point any acquisition at the pack — PGSE, OGSE, real RF/B₁ pulses, any b/Δ/δ, any B₀ and χ, T₁/T₂/ρ, MT. One walk answers questions it was never run for.

    Capability tiers

  • Give a few walkers

    Shards add up and never regress. Contribute from a laptop or a free GPU; sign your shard and you're in the manifest of the solved substrate — a permanent receipt.

    Contributing

The mission

Substrate Commons is a separate thing from the tools that feed it. The engines are edgy, optimised, AI-maintained. The Commons is the opposite: slow, permanent, and open — a shared field anyone can walk into, add to, and take from. Here you have it, friend.