Replicate-aware validation and the identifiability of ingredient attribution in data-driven prediction of industrial rubber compound properties

Compounding plants hold years of formulation records, and it is increasingly common to fit predictive models to them. Those records often contain the same formulation more than once, which quietly changes what cross-validation measures. We analyzed 27 silica-filled natural rubber formulations, each compounded as five independent batches and tested for Shore A hardness, tear strength, tensile strength, and elongation at break. The replicate batches gave the precision of the test program, which combines batch processing and measurement, and with it the 95% critical differences: 0.71 Shore A, 1.33 kN/m, 0.59 MPa, and 29.0%. A random split over the 135 records, which asks how well another batch of a known formulation can be predicted, gave coefficients of determination of 0.90 to 0.98 for tree ensembles. Splitting by formulation, which asks about compositions not yet made, gave at most 0.459 for hardness under grouped five-fold splitting and negative values for the other three properties. The best leave-one-formulation-out estimate for hardness, 0.557 (selection-aware p = 0.0008, conditional on the final candidate set), fell to 0.381 once candidate selection was nested inside the evaluation, and across the four properties the nested prediction error was 3.3 to 7.3 times the critical difference of the test. Silica and polyethylene glycol 4000 were dosed at an almost fixed ratio, so the stable separate Shapley contributions that a random forest gave them are not identified effects. In equal-cost resampling, forty batches spent on twenty formulations measured twice predicted hardness better than the same forty batches spent on eight formulations measured five times. For data of this kind, the number and diversity of distinct compositions, not the nominal record count, limit what a model can be shown to do.