High-fidelity multi-physics simulations are widely used to predict key performance indicators (KPIs) of current interruption in electromechanical circuit breaker (e.g., peak current, arc duration) [1], but their high computational cost limits their use in parametric studies, uncertainty quantification, and operating-margin assessment. This makes surrogate modeling essential, yet challenging because the physics induces a success/failure feasibility split: the current interruption KPIs are meaningful primarily in the successful domain. This regime separation can render standard regression-driven active learning inefficient by oversampling failures and slowing KPI surrogate convergence. We propose a Dual Surrogate Active Learning (DSAL) framework that jointly learns the success/failure boundary and accurate KPI surrogates within the success region, using a simulator-in-the-loop workflow. The method uses a Gaussian Process Classifier [2] for the success probability combined with Gaussian Process Regressors (one per KPI) [2], trained only on successful samples. The acquisition alternates between a boundary query driven by classifier uncertainty and a KPI query driven by aggregated expected prediction error, restricted to candidates predicted to be successful above a confidence threshold. In a case study on an ABB low-voltage miniature circuit breaker, DSAL achieves <5% mean percentage error across all KPIs on a held-out test set using 22 simulations (10 initial plus 12 acquired) while refining the success/failure boundary. Accuracy improves rapidly for challenging quantities. For instance, peak current error decreases from 23.6% to 3.3% within five acquisitions. Active learning operates as intended: boundary-driven queries target the frontier (3 successes/3 failures), whereas regression-driven queries remain mostly feasible (5 of 6 successes), limiting wasted evaluations. DSAL matches grid-level KPI accuracy (64 runs) with only ~25–30% of the simulation budget and achieves test-set classification accuracy of 1.0 versus 0.8667 for the grid sweep. Overall, DSAL provides uncertainty-aware, data-efficient surrogates for fast KPI prediction and operating-margin exploration under feasibility constraints.