Data-Driven Multi-Property Prediction and Optimal Replacement Ratio Discovery for Biomedical Waste Glass Self-Compacting Concrete

Abstract

Healthcare facilities are producing large amounts of biomedical glass wastes (vials, ampoules, and laboratory ware) which are non-biodegradable. Moreover, with the growth in hospital infrastructure, disposal is becoming increasingly complicated. This work is a continuation of an experimental program in which processed biomedical waste glass (BWG) replaces natural fine aggregate in M40-grade self-compacting concrete (SCC) at ten different incremental replacement levels (0–45%, 5% increments). The fresh-state behaviour (in terms of segregation, flowability, passing ability) is evaluated as per EFNARC guidelines and hardened-state compressive, split tensile and flexural strength tests are conducted at 7, 28 and 56 days. The curing age is treated as a second explicit input variable along with replacement percentage. This means that the original ten-mix dataset is converted to a 30-sample matrix (BWG%×age). This matrix is suitable for supervised multi-output regression. A total of four regression models were benchmarked (Random Forest; Gradient Boosting; Support Vector Regression; second-degree polynomial regression) with an 80/20 train-test split using 5-fold cross-validation. The model with the best generalization capability – Gradient Boosting – displayed test R² values of 0.987, 0.951, and 0.980 for compressive, split tensile, and flexural strength respectively. Due to the lack of a similar second feature the ten fresh-property measurements (slump flow L-Box ratio V-funnel time) one Gaussian-noise augmentation scheme (σ = 1.5%) was applied to this subset only and the model outputs checked against the original unaugmented measurements kept as an external validation set (R² of 0.89-0.99) as this as a self-validating activity given its limited experimental basis this component is shown as a supplementary sensitivity check not core evidence. SHAP-based attribution analysis identifies that the clear driver of compressive-strength variation is BWG content while curing age is the clear driver of flexural-strength variation. This provides quantitative support to the micro-filler/pozzolanic mechanism inferred from the raw experimental trends. By using a multi-objective optimization which weights normalized structural desirability against a waste-utilization index, a strict-constraint optimum is found near 20% BWG replacement and a relaxed, IS 456-compliant optimum near 32-33%. These two findings bracket and stretch the raw data-based recommendations of 10-15%. The framework developed provides a data-driven complement to traditional trial-mix design for waste-derived SCC, with the small size of the underlying sample acknowledged as a constraint.The development of sustainable concrete with biomedical waste glass through machine learning and multi-objective optimization.