Thermo-mechanical performance of straw-reinforced adobe bricks with three fiber-length classes: Coupled statistical analysis, two-stage machine-learning prediction, and multi-criteria weighted ranking

Natural-fiber-reinforced adobe bricks offer a lower-processing-energy alternative for building envelopes, but their thermo-mechanical properties depend on mixture composition and the resulting physical state. This study examines 48 straw-reinforced adobe formulations with three fiber-length classes (D1: <2 cm; D2: 2–6 cm; D3: 6–20 cm). Density, porosity, compressive strength (Rc), flexural strength (Rf), and thermal conductivity (λ) were measured. Descriptive statistics and correlation analysis were combined with a physically motivated, two-stage machine-learning workflow: the first stage predicts density and porosity from mixture descriptors; the second estimates Rc, Rf, and λ from formulation variables and physical-state descriptors. Random Forest and Gradient Boosting were assessed using a hold-out partition and reported cross-validation analyses. Across the investigated formulations, increasing straw content was associated with lower density and thermal conductivity, greater porosity, and generally lower compressive strength; mean flexural strength was highest at 1–2% straw. Gradient Boosting performed best overall in the reported internal comparisons. A min–max-normalized weighted score (0.40 for Rc, 0.30 for Rf, and 0.30 for the inverse of λ) ranked B17 (2% straw, D2, 22% water) first among the measured formulations. An illustrative weight-sensitivity analysis showed that this result depends on the chosen priorities: B7 ranks first when the compressive-strength weight increases to 0.60. Cross-validation shows consistency within the available partitions but cannot establish transferability to other soils, fibers, or manufacturing conditions. The approach is therefore intended for exploratory screening and scenario-dependent formulation selection within the studied experimental domain.