Space-filling designs support flexible modeling in noisy multi-response experiments, where some responses may follow simple relationships with factors even when others are complex. An individual design can nevertheless give poor joint precision for low-order effects. We use a low-order working model to guard against such designs while retaining flexibility in the final response analysis. The filter selects one complete candidate from a pool by minimizing geometric loss divided by sphericity, without altering its points. Sphericity measures the balance of the confidence ellipsoid's principal-axis variances independently of its volume. It is the A-criterion's shape factor in the stated coefficient coordinates (raw filtering) and the I-criterion's shape factor in coordinates orthonormal over the prediction region (target filtering). In cube studies with quadratic least-squares and Gaussian-process fits, target filtering reduced aggregate root-mean-square prediction error by 1.4 to 3.3% relative to the geometrically best candidate in the same pool, at a median geometric-loss increase of 0.61%. Its primary benefit is tail protection: in one near-saturated setting, the worst-decile mean working-model prediction standard error fell by about a quarter, at a median geometric-loss increase of 3.2%. In constrained mixtures, raw filtering and A-optimality in physical-proportion Scheffe coordinates can emphasize extrapolated pure-component responses and blending contrasts over regional prediction. Across seven settings, A-optimal designs had 1.3--3.4 times the mean regional working-model prediction variance of I-optimal designs; target filtering improved regional working-model precision in most settings.

翻译：暂无翻译