An independent computational physics lab in Hong Kong. We run fluctuating and multiphase flow simulations, migrate scientific codebases so one source runs on any GPU, and check whether numerical results are determined and hold up. Everything we produce comes with the record that lets someone else re-run it.
Thermal noise is normally the least reproducible part of a simulation.
Ours isn't. Counter-based random number generation inside the metalanguage makes the noise stream a pure function of lattice position and time step — so a fluctuating run produces the same trajectory on CUDA, Metal, OpenCL and CPU. Correct fluctuation–dissipation on a lattice is rare on its own; reproducing it on someone else's machine is rarer.
What we do
Thermal fluctuations done correctly. Getting fluctuation–dissipation right on a lattice means placing noise across the full moment structure, including the non-hydrodynamic sector most schemes leave out.
Nobody ports scientific CUDA for elegance — they port it because they need another backend, and they stall because there is no way to prove the port is faithful.
Moving fp64 → fp32 (or mixed precision) without a measured error budget is a guess.
Whether a computational claim is determined, and whether it holds up — completeness, regeneration with cross-backend agreement, consistency audit. Nothing proprietary has to leave your infrastructure. Not a peer-review certificate for your science; a reproducibility and consistency layer you can defend.
The framework
A kernel is written once in a small metalanguage and lowered to CUDA, OpenCL, Metal or plain C, with the generated source available for inspection at every step. There is no hidden JIT and no graph rewriting: what executes is what you can read.
Device memory, kernel dispatch and synchronisation are reached through one interface, so the same simulation runs on an NVIDIA cluster, an Apple laptop or a CPU-only machine without touching the physics. Cross-backend agreement is a first-class test rather than an aspiration.
The framework is fully open source — the vehicle. Idea.Deploy Labs sells expert-led, gated engagements that use that vehicle on your infrastructure: each unit must reproduce your reference numbers before the next one proceeds. We do not sell a private fork of idea.deploy.
Published work
Not a description of a method — the code that made the numbers, linked from the paper and runnable against the framework.
The lab
People who do computational fluid dynamics rarely arrive from statistical field theory, and statistical physicists rarely write GPU kernels. That intersection is the lab's specialism, and it is what connects work on spin glasses and nucleation to lattice pressure tensors, interface physics and multiphase flow — one thread rather than a scattered publication list.
Founded and directed by Matteo Lulli — PhD in Physics, University of Rome Sapienza — advisor G. Parisi. Previously Research Associate at The Chinese University of Hong Kong, Research Associate Professor at the Southern University of Science and Technology, and postdoctoral researcher at The Hong Kong Polytechnic University. Published in Physical Review Letters, Physical Review X, Physical Review E, Soft Matter and Computer Physics Communications.
Engagement
Most work follows the same shape:
That is the product, not a process deck.
Contact
Prefer a scope call first? Say so in the subject line — we’ll reply with times. Ask for the private offer sheet if you want scope options — no prices on this site.