Surrogate-Accelerated Twins

Surrogate acceleration blends fast approximations with occasional full-physics runs, keeping a twin both quick and trustworthy.

Fast most of the time, exact when it counts

A twin needs speed for control and ensembles but accuracy for trust. A surrogate-accelerated twin resolves the tension by using fast surrogates for the great majority of evaluations while calling the full physics code occasionally, to check the surrogate, to handle inputs outside its trained region, and to generate fresh training data. The result is a twin that is quick in the common case and correct in the demanding one.

The multifidelity idea

Kronos motion — fast proton

Multifidelity methods combine many cheap, approximate evaluations with a few expensive, accurate ones to get the best of both. The cheap model captures the trend; the expensive model corrects the level and the hard cases. A twin arranges its models into such a hierarchy and spends its computation where it buys the most accuracy per unit time.

Knowing when to escalate

Detect when a query lies outside the surrogate's trained domain and defer to full physics
Trigger a full run when residuals suggest the surrogate is drifting
Periodically sample full-physics runs to monitor and retrain the surrogate
Escalate automatically for safety-relevant decisions that warrant the cost

Closing the loop with retraining

Full-physics runs triggered during operation, and smoothed reconstructions of real behavior, become new training data. Retraining the surrogates on this data keeps them accurate as the machine and its operating envelope evolve, and steadily shrinks how often escalation is needed. The twin thus improves itself over its life, under the change-control discipline of governance. See real-time model updating.

In the Kronos twins

Surrogate acceleration is what lets a physically rich twin of the Hyperion breeder or the burner run inside a control cycle and across large ensembles, while retaining the option to fall back to validated full-physics codes for the hard or safety-critical cases. Today both surrogates and full codes run on design models; after the machines operate, real data retrains the surrogates. See surrogate models and real-time computing.