Independent research lab · Compressible & reactive flow
Shocks, detonations, stiff chemistry, refinement that moves with the front. We derive the methods from first principles and implement them ourselves — lean enough to run a two-million-cell airframe on one workstation, portable enough to run the same source on any GPU, and verified to the standard a journal would demand.
An independent research lab
The best numerical methods in compressible and reactive flow are developed inside a small number of large institutions, published on a multi-year cycle, and reach the engineers who need them — if they reach them at all — a decade later, wrapped in a licence. The bottleneck is not the science. It is the gate in front of it.
We are an independent lab built to remove that gate. We keep everything that makes academic research trustworthy — falsifiable claims, canonical verification, cross-code comparison, honest limitations, reproducible artefacts — and we run it on an engineering clock, pointed directly at the problem in front of you.
Accuracy over aesthetics
Our field has learned to sell itself with colour. We optimise for the number being right, and we say so plainly where it is not. The only moving picture on this site is a live solver you can disturb with your cursor — because a computation you can interrogate is evidence, and a rendering is not.
Discipline, engineering clock
Hypotheses stated before testing. Falsification conditions written down first. Canonical verification, cross-code comparison and byte-level reproducibility gating every change. All of it delivered against a project schedule instead of a publication cycle.
An unusual combination
Deriving a scheme from first principles and making it run fast on real hardware are normally two different people in two different departments, trading artefacts across a wall. Here they are the same work — which is why the numerics and the implementation are designed for each other.
What we model
Most CFD is comfortable. These are the flows where the discretisation, the chemistry and the mesh all have to be right at once — and where getting one of them wrong does not lose you a decimal place, it loses you the physics.
01
Transonic through hypersonic, on geometry that came out of CAD rather than a tutorial. Unstructured polyhedral meshes, surface-snapped walls, shock systems resolved by refinement that follows them.
02
Detailed kinetics, stiff source terms integrated in tight coupling with the hydrodynamics rather than split away from it, and an energy formulation that refinement cannot corrupt.
03
Dynamic octree adaptation driven by the solution, running as data-parallel passes over space-filling-curve keys. Nothing goes back to the host to regrid.
2.29M cells
Airframe case on one rank
Mixed hex and polyhedral, with dynamic refinement, inside 16 GB.
217B/cell
Resident footprint
Audited line by line. 241 B/cell with a nine-species reacting registry.
10−13
Conservation across an adapt
Mass and energy closure, checked every refinement cycle.
330sites
Standard-C++ parallel loops
No vendor kernel language anywhere in the solver.
The platform
Three components, built to work as one stack. Each solves a problem that the CFD field has largely chosen to live with.
A header-only, policy-composed compressible and reactive finite-volume solver. Every component — flux, reconstruction, closure, integrator, chemistry — is a compile-time type, fused into one update with nothing dispatched at runtime.
A compiler plugin that takes standard C++ parallel algorithms to a portable GPU intermediate representation, so the same unmodified source runs on any vendor’s hardware — including GPUs that current vendor toolchains have dropped.
A hermetic, content-addressed build and dependency toolchain. One command, no system toolchain assumptions, and a lockfile that regenerates the same binary — which is what makes a published result actually reproducible.
Engineering position
Performance is usually treated as an engineering afterthought in computational physics. We treat it as a scientific variable: a solver that is four times leaner and does not stall on memory bandwidth can afford richer numerics — a better shock sensor, higher-order integration on device, tighter chemistry coupling — that a bandwidth-bound code simply cannot pay for.
Resident state is 217 bytes per cell, audited line by line — roughly four to six times lighter than the mainstream reference codes at the same physics. That is not a micro-optimisation; it is what moves the solver off the memory-bandwidth wall and leaves arithmetic headroom to spend on better numerics.
Block-structured reference codes still orchestrate dynamic regridding and load balancing substantially on the CPU. Refine, coarsen, two-to-one balance and remap are here expressed as sort-and-scan passes over integer-lattice Morton keys — the whole adapt chain is data-parallel, and the host is not in the loop.
Over 330 parallel loops, all written as standard C++ parallel algorithms. No vendor kernel language, no pragma dialect, no second implementation drifting out of sync. The CPU build is the correctness floor and the GPU build is verified bit-identical against it.
Refactors are verified byte-identical on adaptive reacting cases before they land. It is a demanding discipline and it pays for itself immediately: when a result changes, you know exactly which change caused it.
From surface to answer
Not a simplified representation of it. A triangulated surface out of CAD goes in; a partitioned, body-fitted, adaptively refined mesh and a solved flow field come out — through a pipeline we built end to end, so that when the physics needs the mesh to change, it can.
Open questions
We hold these as questions, not as products. Each one states why it is hard, how we are attacking it, and what we have not yet been able to determine.
Capturing a shock costs artificial dissipation. Spend it in the wrong place and you lose the very instability you were trying to predict.
How we approach itChemistry sets a length scale orders of magnitude below the device. Under-resolve it and the computed wave speed is not slightly wrong — it is qualitatively wrong.
How we approach itThe transition is sudden, geometry-dependent, and the single question that matters in hydrogen explosion safety. It is also where every shortcut in a reactive solver shows up.
How we approach itAdaptive refinement is how multiscale problems become affordable. It is also a continuous opportunity to lose conservation, symmetry, or both.
How we approach itHow to work with us
We are an independent lab looking for industry collaborators, research sponsors and academic partners. Three different relationships, one standard of evidence.
Industry collaboration
A shock system your solver smears, a detonation that will not stay stable, a mesh that will not converge, a case that takes a fortnight per run. We take it on as a research problem with an engineering deadline — and you get the method, the validation evidence and the limitations, not a black box.
Research sponsorship
Sponsoring a line of work here buys direct access to results as they happen — not a conference paper in two years. Sponsors shape the problem selection, see the negative results as well as the positive ones, and get the capability applied to their own configurations while it is being built.
Academic partnership
We work the way a good research group works: hypotheses stated before testing, falsification conditions written down first, cross-code comparison, reproducible artefacts. We collaborate with university groups on benchmarks, shared validation and co-authored publications.
Open for 2026
Tell us what you are trying to predict and what has gone wrong so far. If it is outside what we do, we will say so — and point you at who does it.