Explainable-machine-learning-powered discontinuum parametric analyses of the diagonal compression response of unreinforced masonry walls retrofitted with light timber cladding modules

Unreinforced masonry (URM) buildings represent a large portion of the existing building stock worldwide but are highly vulnerable to seismic actions and often exhibit poor thermal performance. To address this dual challenge in Eastern Canada, a cold region exposed to low-to-moderate seismic risk, a light timber modular retrofit system was recently proposed. Building upon previous experimental evidence, this study performs a comprehensive parametric and explainable machine learning investigation to evaluate the diagonal compression behaviour of URM walls retrofitted with this timber system. A three-dimensional Distinct Element Method (DEM) model is developed and calibrated against laboratory tests. The validated model is then employed to generate a virtual experimental database encompassing 576 simulations that systematically vary key mechanical and connection parameters. Statistical interpretation and seven machine learning algorithms are applied to predict the initial and residual peak strength and displacement capacity. All data and models are made openly accessible to support future research and practical implementation. The gradient boosted decision trees, support vector machine, and random forest models achieve the highest overall predictive performance, with an average R 2 score of 0.943, 0.939, and 0.934 over the testing dataset, respectively. SHapley Additive exPlanations are further employed to provide transparent and interpretable insights into machine learning models. The integrated numerical and machine-learning framework established in this study provides a transparent tool for understanding the coupled mechanisms governing retrofitted URM systems and offers insights supporting the future design of sustainable light timber retrofit solutions in terms of diagonal compression behaviour.