Figure 2. Open experimental data and model-structure controls. (a) Published AH36 coupon corrosion rates from Ivanovich et al. [23]; bars show means and error bars show sample SD across four coupons per treatment, nested within two biological replicates. (b) RMSE of sixteen held-out treatment-replicate means, with the two coupons in each unit averaged before scoring. Each biological replicate is held out in turn at all four observed pressure–flow contexts; microbial ratios are calibrated on training controls only. Direct and factorized cell means coincide with their saturated ordinary-least-squares counterparts. Ridge alpha=1 and the XGBoost preset are unchanged; the full-interaction controls distinguish model structure from reconstruction. Data: Ivanovich et al. (2025), doi:10.3389/fmicb.2025.1540664; analysis and figure produced for this study.
