A literature-derived database and machine learning-assisted analysis of compressive strength retention in biochar-modified cementitious composites

Biochar has attracted growing interest as a carbon-rich additive or partial cement replacement in cementitious composites, but published data remain fragmented across feedstocks, processing conditions, mixture designs and material types. This study compiled a structured database from 56 sources and 368 extraction rows. The target variable was strength ratio, defined as the compressive strength of a biochar-modified mixture divided by that of its corresponding control. After excluding 22 control rows that equal unity by definition, the final modelling dataset contained 131 biochar-modified rows: 82 high-confidence Tier 1 rows for the main analysis and 49 Tier 2 rows for sensitivity analysis. Tier 1 random splitting produced R² = 0.335, MAE = 0.118 and RMSE = 0.164, whereas leave-one-source-out (LOSO) validation produced R² = −0.030, MAE = 0.170 and RMSE = 0.207. Statistical testing showed that strength-ratio normalisation substantially reduced, but did not eliminate, material-type effects. Nested group-aware tuning did not restore positive Tier 1 LOSO R² (best R² = −0.008). On 54 independently published experimental mixture means not used for model development, the tuned Tier 1 random forest gave R² = −0.153, MAE = 0.082 and RMSE = 0.108; the strong rank correlation (Spearman ρ = 0.797) but positive prediction bias indicated partial trend recognition without reliable calibration. Missing-data sensitivity, source-level Friedman and Wilcoxon-Holm tests, and grouped SHAP analyses reinforced the same interpretation. Biochar dosage remained the leading predictor, but the framework is best understood as exploratory, diagnostic and source-aware rather than as a universal or deployable prediction model.