Prediction and optimization of the thermomechanical performance of carbon-free Adobe bricks reinforced with straw and sawdust using machine learning

Abstract

This study proposes and applies a machine-learning-driven optimization framework to predict and enhance the thermomechanical performance of carbon-free adobe bricks reinforced with straw and sawdust. To move beyond trial-and-error mix design under a strength–insulation trade-off, the study establishes reproducible mix-selection rules that reduce experimental iterations. Experimental tests show that adding small amounts of straw (1% and 2%) significantly improves compressive strength, increasing it from 5.41 MPa to 9.62 MPa (+78%) and 7.93 MPa (+46.5%), respectively; however, higher dosages lead to a decrease in strength due to excessive porosity. Sawdust reduces mechanical strength but improves insulation by lowering thermal conductivity from 0.632 W/m.K for the reference brick to 0.145 W/m.K at 10% sawdust. Mixed formulations provided the best compromise: with approximately 0.5–4% sawdust and 0.5–4% straw, they maintained compressive strengths above the minimum requirement of 2.07 MPa established by the Mexican adobe construction standard. A measured dataset (density/porosity, Rc/Rf, λ and Cp) was used to train surrogate models with a 70/15/15 train–validation–test split, 5-fold cross-validation, and grid-search tuning. The machine learning models exhibited distinct predictive capabilities, achieving R² = 0.323–0.566 for compressive strength and R² = 0.794–0.991 for thermal conductivity, and multi-objective optimization (Pareto-based selection) further revealed that hybrid mixtures offer the most balanced solutions. These findings confirm the potential of agricultural waste valorization for the production of eco-friendly building materials and establish a systematic methodology that combines experimental work with artificial intelligence to optimize sustainable adobe bricks.