Surrogate Acceleration of the Twin
Full multi-physics solves are too slow for real-time control, so Kronos distills them into fast surrogates that preserve the physics that matters.
The speed requirement
A first-principles solve, Monte Carlo neutron transport, finite-element equilibrium and thermomechanics, ideal-MHD eigenvalue analysis, can take minutes to hours. The predictive shadow needs a full coupled state every few tens of milliseconds. The gap is many orders of magnitude, and no amount of hardware closes it directly. Kronos closes it by replacing the online solve with surrogates, models trained offline to reproduce the expensive solver's output at a fraction of the cost.
Kinds of surrogate Kronos uses
- Physics-informed networks: equilibrium, stability, ambipolar potential (physics in the loss)
- Neural operators: learn solution operators for parametric transport/thermal fields
- Reduced-order models: project high-dimensional fields onto dominant modes
- Distilled Monte Carlo: fast neutronics surrogates trained on offline transport runs
The choice per module depends on the physics. Where a governing PDE is known and enforceable, PINNs give physics-grounded surrogates. Where the map is from parameters to full fields (neutron flux over the blanket, temperature over the divertor), neural operators learn the solution operator and generalize across parameters. Reduced-order models exploit that these fields live near a low-dimensional manifold.
Offline cost, online speed
The expense moves offline to L0: Monte Carlo campaigns, PINN training, and surrogate distillation run on the GPU clusters and batch-retraining pipeline. The runtime twin only evaluates the trained surrogates, at fixed, low cost. This offline/online split is the enabling idea behind the whole L3 stack, and it is why the shadow can be both fast and physically faithful.
Surrogates never replace the reference solvers; they stand in for them online. The full solvers remain the validation oracles and the fallback for out-of-distribution queries, so speed is bought without abandoning first-principles ground truth.