Propulsion is no longer a drafting problem. It is a compute problem.
Engines are usually drawn by hand in CAD, simulated for weeks and redrawn. We want to turn that around: state the mission, let physics-informed AI search the design space, and get a part that can go straight to a metal 3D printer. The same approach should later shape magnetic fields for plasma drives and, one day, spacetime metrics.
How WarpCore would work
01
Mission to math
No sketches, only constraints: thrust, specific impulse, propellant, available volume, thermal limits and budget become boundary conditions.
02
Physics-informed AI
Neural networks that carry the governing equations (Navier–Stokes, heat transfer, Maxwell) in their loss, or learn whole solution operators, predict a new design in milliseconds instead of a CFD run over days.
03
Geometry as code
Shapes are generated from equations and voxels, not drawn. Injector, chamber and cooling channels can be one weld-free part, with channels routed where the wall runs hottest.
04
Direct to metal
The output is a print-ready file for laser powder bed fusion in copper alloys such as CuCrZr or nickel superalloys such as Inconel 718.
The AI models
Generic AI cannot design an engine: a language or image model does not know that gas must obey conservation of mass, momentum and energy. WarpCore combines several kinds of models, each with a clear job.
Learns the physics
Physics-informed neural networks (PINNs)
The governing equations, Navier–Stokes for flow, heat conduction and Maxwell for magnetic fields, are written into the network’s loss function. The network is penalised whenever its prediction breaks the physics, so it needs far less training data and does not simply guess.
Raissi, Perdikaris & Karniadakis, 2017
Predicts in milliseconds
Neural operators (Fourier Neural Operator)
Instead of solving one case, a neural operator learns the mapping from a shape and its boundary conditions to the whole flow, pressure and temperature field. Once trained, it evaluates a new nozzle in milliseconds instead of a CFD run over hours or days.
Li et al., 2020
Trained on real simulations
CFD-trained surrogate model
The first 100 to 500 designs are simulated conventionally with OpenFOAM, using combustion data from NASA CEA. These results train and calibrate the fast models, so their predictions stay anchored to established solvers.
OpenFOAM · NASA CEA · NVIDIA PhysicsNeMo
Searches the design space
Reinforcement learning and evolutionary search
An optimiser proposes, mutates and scores tens of thousands of geometries per hour against the mission targets: thrust, efficiency, wall temperature and mass. The fast surrogate makes this search affordable.
PyTorch
Horizon 2
Multi-agent reinforcement learning for plasma
For magnetic nozzles and plasma thrusters, several cooperating agents shape coil positions and field topologies that are too non-linear to tune by hand, with stability as a constraint.
Planned research
Horizon 3
Neural models for Einstein’s equations
The same idea applied to general relativity: networks that search for warp geometries with low, positive energy, each candidate checked against the energy conditions with Warp Factory.
Helmerich, Fuchs, Bobrick et al., 2024
The open-source toolchain
Every layer of the pipeline already exists as open software. Our work is to connect them and train the AI layer in between.
Numerical toolkit that checks warp-drive spacetimes against the energy conditions
Proof that the idea works
Computational engineering of rocket engines is not science fiction. LEAP 71’s Noyron designs rocket engines algorithmically, and engines designed this way have been 3D-printed and hot-fired. Most of that work targets chemical engines. We want to bring the approach to in-space propulsion, where electric and plasma thrusters depend on magnetic fields that are hard to optimise by hand.
Three horizons
Horizon 1
Chemical and high-power electric
Generative design of orbit-manoeuvring thrusters, green monopropellant engines and Hall or magnetoplasmadynamic channel geometries for satellites and orbital transfer vehicles.
Horizon 2
Plasma and fusion confinement
Training the same models on magnetic nozzles and plasma stability, the problems behind our HELIOS and ECHO concepts.
Horizon 3
Spacetime metric engineering
Extending the models to Einstein’s field equations to search numerically for slower-than-light, positive-energy warp geometries, connecting to tools like Warp Factory.
How we develop it, step by step
We start small: WD-01, a 100-newton satellite thruster with regenerative cooling, running on monopropellant or methane and liquid oxygen. It is small enough to design on one workstation and touches every part of the pipeline.
Physics baseline
Compute combustion temperature, pressure and exhaust properties for the chosen propellant and size the nozzle with classical rocket equations. This gives the reference every AI result is compared against.
NASA CEA
Parametric geometry
Describe the engine as code: the Laval contour from chamber through throat to exit is set by five parameters, and the cooling channels spiral through the wall, moving closer to the inner surface where the heat load is highest.
PicoGK · voxels and signed distance fields
Training data
Simulate the first 100 to 500 variants with full CFD to obtain flow, pressure and wall temperatures.
OpenFOAM
Train the surrogate
Train a physics-informed network or neural operator on these results, with the governing equations in the loss, until it predicts unseen designs within a set error of the CFD.
NVIDIA PhysicsNeMo · PyTorch
Optimise
Let the optimiser explore tens of thousands of geometries against the mission targets and keep the best candidates.
Reinforcement learning · evolutionary search
Verify
Re-simulate the best designs with full CFD. Only designs that the conventional solver confirms move on; the discrepancies are fed back to improve the model.
OpenFOAM · ParaView
Print and publish
Export the winning part for laser powder bed fusion, render a cross-section and publish the results and lessons on this site. Hot-fire testing would follow with a test-stand partner.
STL export · Blender
Hardware to start: one current NVIDIA GPU, local or in the cloud.