Machine Learning Surrogates for Unreinforced Masonry Tensile-Strength Prediction

With advances in computational modeling, finite- or discrete-element method-based computational models are often used to conduct nonlinear structural analysis of masonry construction. However, such detailed models are computationally intensive, limiting their usefulness in preliminary analysis or applications requiring real-time simulations (e.g., developing digital twins). A possible solution, increasingly used in other building science areas (e.g., energy efficiency), is developing machine-learning-based surrogate models that mimic the performance of complex physics-based simulation models at significantly reduced computational costs. However, little is known about the premise of this approach in the context of masonry tensile strength prediction due to the scarcity of such applications in the literature. This research proposes a framework to develop and evaluate the performance of machine learning surrogate models in emulating the performance of masonry tensile-strength prediction models. A five-step methodology is proposed: (1) develop computational physics-based models based on the discrete element method (DEM), (2) validate the proposed computational models, (3) generate a data set through parametric variations to support surrogate modeling, (4) train and test data-driven surrogate models to emulate the capabilities of the computational models, and (5) conduct a sensitivity analysis to determine the most influential input parameters. The presented generic approach is demonstrated using a validated discontinuum-based computational modeling strategy based on the DEM. The model predicts the tensile strength and corresponding strain value of unreinforced masonry walls subjected to diagonal compression forces. Results show that the gradient boosting (GB) surrogate models consistently achieved high accuracy levels (R2>0.9) even when using small data sets (e.g., 100 samples). However, compared to linear regression, GB exhibited increases in training and tuning time for large sample sizes. The proposed framework offers unique insights into the premise of data-driven surrogate models to complement and support computational-based techniques, balancing predictive accuracy and model complexity. Additional case studies are required to generalize results to other masonry configurations and contexts.