Train machines
on real physics.

Models train directly on the live GPU simulation through our CUDA bridge, with WL3 hypergraph metrics computed by the solver itself during training. The solver physics runs alongside inference, self healing their predictions, as in our bucket excavator agent.

The Bucket Excavator Agent ↓ Request an evaluation

Multiphysics on one GPU: coupled SPH viscous melt and cohesive DEM particles in a twin screw extruder, the kind of physics a learned model can train on.

The CUDA bridge

The solver library runs inside the Python process. Models train on the live GPU simulation and run beside it at inference, with no copies and no files between them.

Live GPU tensors

CUDA device pointers to positions, orientations, velocities, forces, polyhedron state and Element stress are used as PyTorch and CuPy tensors in place, so models train on the running simulation.

WL3 hypergraph metrics

The solver itself computes WL3 hypergraph metrics of the full contact network during training; the high fidelity data trains a hypergraph neural network that learns how the process itself works, such as mixing.

Self healing inference

The solver physics runs alongside inference: networks and controllers act in the middle of a time step, and Jacobi relaxation pushes their predictions back to physical states on the GPU.

Training data

Training frames with the full contact network to HDF5, or live coupling with no files, for AI physics surrogates, robotics training, reinforcement learning and digital twins.

Hopper discharge: accurate DEM, DEM with ML in the loop, simple DEM — **ML in the loop** (middle) reproduces the accurate DEM hopper flow (left); simple DEM on the right.

The bucket excavator agent

A learned controller trained on calibrated GPU DEM runs infers the hidden particle state from the drive signals and checks its predictions against a live DEM simulation.

1 Ground-truth physics

The DEM of the buckets is calibrated against the real machine. Cohesion cases (clay binder, freezing) build a database of sensor and motor signals to DEM states: contacts, payload, force chains, rock fragments.

2 Live co-simulation

Pre-trained on recorded DEM runs, then the agent acts in closed loop: speed change, jog, reverse. Blaze returns the physical response and the agent learns the causal effect.

3 Read the machine

During operation the agent infers force chains, payload and boulder obstacles from the drive signals alone: torque, tension, power and speed.

4 Check before acting

Predicted high tension states are checked against a live DEM simulation, avoiding high-tension events and improving fill per pass.

DEM of the buckets coloured by force with the live machine signals — DEM of the buckets coloured by force, with the live machine signals.

Learned torque signature wave — A learned torque signature.

DEM verification constructed from the start state force chains — DEM verification built from the start state force chains.

Classify, predict, intervene

The agent compares short windows of drive torque with the learned torque signature library and classifies the machine state:

Free flow: steady transport with a regular torque oscillation, high fill and no carryback.
Cohesive build-up: material sticking to the buckets raises the baseline belt tension and damps the peaks.
Active jam: local particle bridging gives a torque and tension spike that trips the machine if untreated.
Post-jog recovery: the agent jogs and briefly slows the feed, and the flow recovers.

Free-flowing ore: bucket forces and torque signatures streamed to the agent.

Moist cohesive ore: JKR surface energy and liquid bridging give hang-up, boot packing and carryback.

Train machines on real physics.

The CUDA bridge

The bucket excavator agent

Classify, predict, intervene

Train machines
on real physics.