Systematic accuracy of surrogate actuator gradients

Systematically characterize the accuracy of the machine-learning surrogate’s actuator-gradient estimates used by the ARC V3A quadratic-program controller.

Background

The controller obtains growth-rate sensitivities by differentiating a neural-network surrogate with respect to PF-coil currents. The paper reports discrepancies between surrogate partial derivatives and finite-difference total derivatives, especially for coils that also affect plasma shape, making systematic gradient validation an unresolved problem.

References

The principal open questions are robustness to diagnostic noise and reconstruction errors, systematic characterization of surrogate gradient accuracy, generalization of the surrogate beyond the training distribution, real-time implementation on PCS hardware with deterministic timing guarantees, and design of a fallback hierarchy for safe degradation.

— On the feasibility of model-based feedback control of vertical instability growth rate using out-vessel coils in ARC-like scenarios  (2608.22677 - Kumar et al., 24 Aug 2026) in Section 6, Conclusions

The principal open questions are robustness to diagnostic noise and reconstruction errors, systematic characterization of surrogate gradient accuracy, generalization of the surrogate beyond the training distribution, real-time implementation on PCS hardware with deterministic timing guarantees, and design of a fallback hierarchy for safe degradation.

— On the feasibility of model-based feedback control of vertical instability growth rate using out-vessel coils in ARC-like scenarios  (2608.22677 - Kumar et al., 24 Aug 2026) in Section 6, Conclusions