Plasma Copilot: Scenario Optimization

Bayesian and surrogate-accelerated search over the copilot's waveform parameters, using the twin as the objective and the safe envelope as a hard constraint.

From a draft to an optimized operating point

Once the Plasma Copilot has drafted a scenario family it optimizes the free parameters — ramp rates, flat-top setpoints, fueling schedule, shaping coil currents — against a twin-evaluated objective. Because a full twin evaluation is expensive, the copilot drives the search with surrogate models and reserves high-fidelity twin runs for candidates the surrogate flags as promising.

Bayesian optimization loop

The copilot uses Bayesian optimization with a Gaussian-process surrogate over the objective. It proposes the next parameter vector by maximizing an acquisition function (expected improvement) subject to a feasibility model that predicts envelope membership, then evaluates the true objective on the twin and updates the posterior.

# Bayesian optimization with a feasibility constraint
while budget_remaining:
    x_next = argmax_x  EI(x | GP_obj) * P_feasible(x | GP_con)
    if not envelope_check(x_next):        # L4 hard constraint
        GP_con.update(x_next, feasible=False); continue
    y = twin.evaluate(x_next)             # tau_E, betaN, disruption risk
    GP_obj.update(x_next, y)
    GP_con.update(x_next, feasible=True)
x_star = best_feasible(GP_obj)

Objectives per machine

• Breeder: maximize confinement time / stability margin at fixed 9.66 MA flat-top and delta -0.30
• Breeder: frame tritium-breeding scenarios across the 1.1 / 1.5 / 1.8 TBR lever as context, not a tuned setpoint
• Burner: maximize direct-energy-conversion window width at the 26.49 T plug operating line
• Burner: hold end-plug density and ambipolar potential inside the confinement-stable band

Optimization never overrides the envelope. Infeasible candidates are recorded to sharpen the feasibility model; they are not clipped into the feasible region and silently proposed. The optimizer returns a Pareto set (for example confinement versus heating effort) so the operator chooses the trade-off. All runs are logged for audit and provenance, and the surrogate/twin agreement is monitored so the copilot falls back to full twin evaluation when the surrogate drifts.

The optimized point is a proposal. Execution still requires L4 authorization and human approval, and the resulting operating point is handed to disruption advising for continuous margin monitoring during any real shot.