When a screened pattern search spends more simulator calls: evidence from mineral-processing circuit cases

Surrogate models can reduce the cost of optimizing computationally expensive process simulators, but their value depends on whether screening eliminates enough exact evaluations to offset the evaluations it adds and whether it changes the selected operating decision. We compare a Gaussian-process-screened generalized pattern search with the same progressive-barrier search without screening. Both procedures use the same starts, bounds, objective weights and exact circuit evaluator. In 72 target-mode variants spanning twelve authored mineral-processing cases, the screened procedure used 26,367 exact engine evaluations, compared with 22,487 without screening, a 17.25% increase. It used more evaluations in 68 variants, the same number in four and fewer in none. Both procedures returned feasible optima in 70 variants and infeasibility in two; their terminal outcomes matched in 65. In the seven remaining feasible comparisons, absolute differences in recovered primary metal did not exceed 0.0027 t/h (0.017%). Across 5,811 surrogate-selected proposals, 901 improved the incumbent. Absolute recovery-prediction error had a mean of 0.71 percentage points, a median of 0.48 points and a maximum of 38.16 points. The screen therefore preserved outcomes closely on this authored suite but did not reduce simulator cost. The results do not establish performance on plant-calibrated circuits.Source code, data and computational records: https://github.com/fsantibanezleal/CAOS_OreFlow