Prediction of waste shear strength parameters in open dumps using electrical resistivity tomography and composition data through machine learning approaches.

The assessment of municipal solid waste (MSW) stability in open dumps presents significant challenges due to heterogeneity, biodegradation processes, and limitations of conventional invasive characterization methods that yield discontinuous data and disrupt operations. This research developed artificial intelligence models to predict waste shear strength parameters by integrating waste composition data with non-invasive electrical resistivity tomography (ERT) measurements. Large-scale direct shear testing was conducted on 66 waste specimens from fresh and young-age waste zones at an open dump in central Thailand. Waste components were categorized into four groups based on electrical conductivity and mechanical properties: L1 (conductive materials reducing shear strength), L2 (non-conductive materials reducing shear strength), H1 (conductive materials increasing shear strength), and H2 (materials enhancing both conductivity and strength). Three-dimensional electrical resistivity measurements were performed using Schlumberger array configuration with 36 electrodes at 8 cm spacing. Multiple machine learning algorithms were evaluated, with an attention-enhanced neural network using AdamW optimization demonstrating superior performance, achieving Mean Absolute Percentage Error (MAPE) of 6.81 % with R2 = 0.95 for cohesion and 2.98 % with R2 = 0.95 for friction angle. This study advances current practice by integrating laboratory-calibrated ERT features with waste composition in an attention-enhanced architecture, addressing limitations where resistivity alone cannot distinguish physical mechanisms while composition alone cannot capture spatial variability. Shapley Additive exPlanations (SHAP) analysis revealed that density and waste composition exhibit complex non-linear relationships with shear strength parameters, with electrical resistivity serving as an integrated proxy for moisture content, density, and material composition effects. The findings demonstrate feasibility of integrating non-invasive geophysical methods with artificial intelligence for accurate waste characterization. However, limited dataset size (n = 66), laboratory-scale measurements, and site-specific validation restrict model generalizability, requiring future field-scale calibration and multi-site validation.