Frontier · available to early teams
The models.
Frontier turns the simulations an organisation already owns into physics models it can trust and deploy, for any of the physics above. It is where the evidence lives.
See FrontierEngineering physics,
at the speed of design.
DimensionLab builds software for the engineers designing vehicles, aircraft, reactors, power grids, the materials inside them and the molecules inside us.
Mission
Put high-fidelity physics
inside every design decision.
Simulation is how engineers test what they cannot yet build. It is also slow, so most designs are never simulated at all. Models trained on simulations can answer in seconds instead of hours, but only when they learn from enough of the right data, and only when their answers can be checked.
What the research says · Ashton, Brandstetter, Mishra, Fluid Intelligence, arXiv 2511.20455
- 01
The data is the expensive part.
For a 3-billion-parameter CFD foundation model trained on 200,000 high-fidelity simulations, the authors estimate about $10.3M of compute to generate the data and $0.4M to train the model. Every simulation a company has already run is worth keeping.
Table 4 - 02
Simulations are not interchangeable.
“Unlike LLMs, CFD surrogates require an explicit parameter vector to define the properties of the underlying simulation.” Physics model, mesh and numerics have to travel with the data, or pooling it makes models worse.
§3.2, §4.1 - 03
The solvers have to change too.
Foundation models for fluids are “inextricably linked to the development of a new generation of simulation tools”: GPU-native solvers, co-designed with machine learning, that generate high-fidelity data quickly and cheaply.
§8
So we work on both sides of the loop. Frontier keeps every simulation, its setup and its provenance attached to the models trained on it, and measures what those models get right and wrong. And we are building the faster, verified simulation the next generation of models will need, for far more than fluids.
Vision
Every frontier
is a physics problem.
Fusion plants, engineered cells, cargo landing on Mars, grids that never fail, materials designed atom by atom. Each is decided by simulations that take days on supercomputers today. We want them to take seconds, and we want every engineer to be able to trust the answer.
Live · Magnetic confinementCharged particles in a tokamak field: toroidal B ∝ 1/R with a twisted poloidal component, q from 1 at the axis to 3 at the edge; ions and electrons at a reduced mass ratio
01Fusion and quantum energy
Nuclear, plasma and electrical engineering
- Magnetohydrodynamicsplasma stability and confinement
- Particle-in-cellthe kinetics of individual particles
- Quantum Monte Carlosubatomic interactions
Ions (orange) and electrons (blue) gyrate along helical field lines. Without the twist, the ∇B drift would empty the torus in seconds; with it, drifts close on themselves. Confinement is geometry.
Live · A chain finds its foldA 96-bead hydrophobic–polar chain in explicit solvent, periodic box, bath temperature cycling between 0.5 and 1.6 kT/ε
02Bio-digital engineering and synthetic biology
Biomedical, chemical and genetic engineering
- Molecular dynamicsprotein folding and drug binding
- Agent-based modellingthe behaviour of cells
- Microfluidic CFDlab-on-a-chip devices
Hydrophobic beads (orange) avoid the solvent and bury themselves in a core as the bath cools; heat it and the chain unfolds again. The same thermodynamics decides whether a drug binds.
Live · Mars entry, belly firstMach 6 freestream of CO₂ (γ = 1.3) around a Starship-class body at 65° angle of attack, started from undisturbed flow; 3-D on desktop GPUs, 2-D on phones
03Planetary engineering and deep space
Aerospace, materials and terraforming engineering
- Hypersonic CFDatmospheric entry at Mach 5 and beyond
- Discrete element methodlunar and Martian regolith
- Thermal-structural FEAradiation and thermal stress
A Starship-class vehicle enters the Martian atmosphere at Mach 6 with its belly to the flow. The bow shock does the braking and holds the hottest gas away from the skin; the angle of attack is the heat shield. Every cargo landing starts with this calculation.
Recorded · A weir at its design floodHydropower weir from an engineering study at its design discharge of 566 m³/s: upstream level 51.9 m, tailwater 45.25 m, the measured inlet velocity profile, a pressure outlet downstream, the water surface tracked cell by cell; 425 million cells at 10 cm, 20 s of spin-up, 40 s recorded
04Next-generation grids and autonomous robotics
Electrical, hydraulic, robotic and systems engineering
- Free-surface CFDhydropower weirs, spillways and intakes
- Physics-informed neural networksreal-time physical proxies
- Digital twinsliving replicas fed by sensors
- Stochastic optimisationbalancing the load on a grid
Water drops 6.6 m over the crest at 566 m³/s. The nappe, the hydraulic jump in the stilling basin and the surge against the piers decide how the structure wears and how much energy the plant can take. 425 million cells, computed in 71 minutes on one GPU and compared with the plant's conventional CFD study.
Live · Light inside a band gapSquare lattice of dielectric rods (εr = 8.9, r = 0.2a) with a missing row that turns a corner; continuous source at a/λ = 0.36, absorbing edges
05Metamaterials and quantum computing hardware
Materials science and solid-state engineering
- Density functional theorynew crystals from quantum mechanics
- Finite-difference time-domainlight meeting matter in photonics
The lattice of dielectric rods has no modes at this frequency, so light cannot enter the crystal at all. It can only follow the missing row of rods, and it turns the corner with it.
What has to change
Three shifts
already under way.
- 01
From isolated solvers to coupled systems.
A fusion reactor cannot be designed one physics at a time: the heat changes the structure of the coils, the coils shape the field, and the field holds the plasma. The same is true across scales, from the atoms of a material to the stress in a part. Simulation has to become multiphysics and multi-scale, and so do the tools that run it.
- 02
From solving to learning.
Solving Navier–Stokes or Maxwell for a new design can take days on a supercomputer. Physics AI, from PINNs and Fourier neural operators to DeepONets, MeshGraphNets and Transolvers, learns from those solutions and answers in seconds, with the conservation laws built into how it is trained. A model is only as good as the simulations it learned from, and the evidence that it was checked.
- 03
From models to twins.
A digital twin is not a 3D file. It is a replica of a turbine, a grid or a heart that updates with the sensors on the real thing, so engineers can ask what happens next and see a failure weeks before it would happen. It needs fast physics, trustworthy models and a place where the data stays connected.
Where this goes
Second product · in stealth
The solvers.
Frontier learns from simulations; our next product makes them. We are building a new generation of GPU-native solvers for the physics of every frontier here: fast, validated against reference cases and optimised to the last kernel, so the datasets that large physics models learn from can be generated at will, in any domain.
How we build
The agents.
Inside DimensionLab, swarms of AI agents are taking over the engineering: they operate Frontier, write and verify solver kernels, run 500+ verifications and review each other’s work, while people set the goals and check the evidence. We build our products so that the same swarms can operate them for us as well as for you. That is how a small team takes on fusion, cells, planets and grids at once, and how humanity gets more engineering done than it has people for.
What we build
Physics AI platform
Frontier
Turn simulation archives into physics models. Prepare data, train and evaluate surrogates, and deploy them on your own infrastructure, with the evidence attached. Measured on DrivAerML at 0.84% mean drag error across 48 held-out car designs.
Second product · in development
The other side of the loop.
Our next product works on the simulations themselves: a separate product, built for every physics in our vision. It is not public yet. When it is, every claim about its speed and accuracy will come with a validation report you can reproduce.
How we work
Three rules
we do not bend.
- 01
Measured, not claimed.
We publish the numbers behind our results with the data, hashes and method needed to reproduce them. A claim about speed or accuracy without that is a promise, not a result.
- 02
Your data stays where it is.
Our software runs inside your infrastructure: a workstation, your cluster, a private cloud or an air-gapped site. Simulation archives and trained models never have to leave.
- 03
Open where it connects.
Formats, connectors and data libraries are Apache-2.0 on GitHub, so the work you build on them is never locked to us.
Browse the open ecosystem
From the lab
Contact
Working on something
that has to be right?
Tell us what you are simulating and what is slowing you down.
Or