Prediction of mechanical properties of Colombian adobe masonry using machine learning and experimental correlations

Adobe masonry structures exhibit considerable variability in their mechanical properties and high vulnerability to seismic loading. These characteristics hinder the development of reliable predictive models and design criteria, particularly in countries such as Colombia, where locally produced adobe differs significantly from the materials considered in international standards. This study develops models for predicting the mechanical properties of adobe masonry in Colombia using both simplified empirical correlations and machine learning approaches. A comprehensive database consisting of 344 experimental results was compiled, including measurements of compressive strength of adobe units, flexural strength of adobe units, compressive strength of masonry, elasticity modulus, shear strength, and shear modulus. Simplified equations calibrated for Colombian adobe were proposed following the conceptual framework of the NZS 4297 standard. In addition, artificial neural networks and extreme gradient boosting (XGBoost) models were implemented to predict the compressive strength and elasticity modulus of adobe masonry using relevant three key input variables: adobe unit strength, wall aspect ratio, and mortar type. Machine learning models were trained and validated on the subsets of masonry-level data available for each property ( N = 34 for compressive strength, N = 38 for elastic modulus) using 5-fold cross-validation to assess model robustness given the limited sample size. Model performance was evaluated using multiple statistical indicators, including correlation coefficient, root mean square error, mean absolute error, and mean absolute percentage error. Results indicate that machine learning models significantly outperform simplified empirical equations. In particular, XGBoost achieved the highest predictive accuracy on the independent test subset, with correlation coefficients of 0.96 and 0.84, RMSE of 0.08 MPa and 18.36 MPa, and MAPE of 7.71% and 35.01%, for compressive strength and elasticity modulus, respectively. An interactive MATLAB-based platform was also developed to facilitate the practical application of the proposed models in the analysis and design of adobe structures.