Resource-efficient machine learning via multi-staged feature elimination for sustainable structural assessment and life-extension of FRP-strengthened RC beams

Retrofitting existing structures to extend service life is vital for reducing the built environment's massive footprint, which accounts for 40% of global energy use and solid waste. Fiber-Reinforced Polymer (FRP) strengthening is a widely used retrofit technique for reinforced concrete (RC) beams. However, assessing whether a strengthened beam will perform safely traditionally depends on extensive laboratory testing of geometric and material properties, a process that is costly, slow, and resource-intensive. This study evaluates whether reliable structural assessment can be achieved with a reduced parameter set, thereby minimizing the testing burden. Using an experimental database of 624 FRP-strengthened RC beam specimens, we developed machine learning models that jointly predict governing failure mode and ultimate load capacity, achieving Macro-F1 = 0.93 and MCC = 0.91 for classification, and R 2 = 0.96 (RMSE = 9.59 kN) for capacity prediction under a random specimen-level evaluation protocol, with additional grouped and leave-one-study-out cross-validation performed to assess generalization across independent source studies. Beyond predictive accuracy, this study's central contribution is a staged feature-elimination pipeline combining Boruta screening, cross-validated SHAP and permutation importance ranking, and factorial ablation with paired statistical significance testing corrected for the non-independence of repeated cross-validation folds and for multiple comparisons. Results demonstrate that two input variables, including FRP tensile strength, whose determination requires destructive coupon testing, can be eliminated with no statistically significant loss in predictive performance on the holdout test set, a conclusion further supported by robustness checks across alternative model families and by uncertainty quantification via bootstrap confidence intervals. This reduces the required input set from 12 to 10, establishing a validated minimum viable assessment parameter set that meaningfully cuts the testing, time, and resource burden of retrofit decision-making, while also lowering model training time by approximately 15%, in qualitative alignment with SDG 9, SDG 12, and SDG 13.