AI surrogates for physics simulation

Simulate in seconds what used to take weeks.

Our models run 10⁴–10⁶× faster than the codes used in fusion R&D: particle-in-cell, gyrokinetic and MHD. Faster simulation means more design iterations, and more iterations means a shorter path to fusion. We built them on the hardest case there is: turbulent plasma, outside the fluid regime, and they learn it from a minimum of training simulations.

Talk to us How it works

Evaluation time

Illustrative · logarithmic axis

Calda surrogatesingle GPU

—

Plasma simulation codeparticle-in-cell · GKW · MHD

0 ms

One reactor-relevant run can cost millions of CPU-hours. The surrogate answers in milliseconds.

Time to evaluate one operating point, on a logarithmic axis. Speedups of 10⁴–10⁶× are measured against standard particle-in-cell codes; over 10⁴× has been demonstrated on real plasma simulations. Chart is illustrative; benchmark data available under NDA.

The bottleneck

Fusion moves at the speed of its simulations.

Standard plasma codes, such as WarpX and other particle-in-cell solvers, are physically faithful but extremely expensive. Gyrokinetic and MHD codes carry the same cost. A single reactor-relevant run can cost millions of CPU-hours.

Design iteration, control synthesis and uncertainty quantification become impractical. The constraint on fusion R&D is throughput.

Our models learn from those codes and stand in for them. Same physics, orders of magnitude less compute, and so orders of magnitude faster.

Faster answers mean more design iterations. More iterations mean a shorter path to fusion.

Architecture

Fast, without losing the physics.

Resolves the small scales

The local component captures small-scale structure that other AI surrogates lose. Fine-scale physics survives the speedup.

Grid-free

Plasma confinement isn't cubic. Not being bound to a Cartesian grid means realistic reactor geometries, and far better generalisation across configurations.

Stable training

A physics-informed parameterisation gives stable training dynamics, and realistic evolution over long horizons.

Few training runs

Because the physics is built into the architecture, our models learn from a minimum of training simulations. You do not need to generate a large dataset before the model is useful.

Speed-up: 10⁴–10⁶×
Benchmarked against: particle-in-cell · GKW · MHD
Training data needed: minimal
Generalisation: transfer across configurations
Deployment: on-prem · cloud · API
Your data: never used to train for others

Over 10⁴× demonstrated on real plasma simulations, not internal benchmarks of our own design. Your simulation data is never used to train models for anyone else.

Beyond fusion

Plasma today. Fluids next.

Fusion is the focus today. The same architecture applies wherever the simulation is the bottleneck.

Fusion & plasma

Kinetic, turbulent plasma outside the fluid regime: the hardest case, and where we started.

Fluid dynamics

Aerodynamics and industrial fluid dynamics, where design iteration runs into the same compute wall.

Visual effects

Fluid and plasma simulation for film and games, where the constraint is turnaround rather than tolerance.

Core team

Six PhDs in physics, AI and statistical inference.

Ben Wandelt, PhD

Co-founder

Bloomberg Distinguished Professor, Johns Hopkins University.

Niall Jeffrey, PhD

Co-founder

Faculty, King's College London. AI Institute and Physics.

Justin Alsing, PhD

Co-founder

Former faculty, Stockholm University.

Lucas Makinen, PhD

Science

Physics-AI Fellow, University of Cambridge.

Mona Wilcke, PhD

Commercial

Commercialisation of scientific innovation.

Hamza Bokhari, PhD

Head of Business Development

Commercial strategy, pilots and industrial partnerships.

Award-winning researchers in physics and AI, with a decade of work on extracting physics from simulations too expensive to run at scale.

Contact

Waiting days for a simulation?

Tell us which code you run and which decision you're trying to make faster. If a surrogate fits your problem we'll say so. If it doesn't, we'll say that too.

hello@calda.ai LinkedIn