Data-Driven Machine Learning Surrogates to CTHBM Model for MSW Landfill Settlement Prediction

Municipal solid waste (MSW) landfills are often characterized by significant post-closure settlements that occur over extended periods. These settlements primarily occur due to the overburden stresses, biodegradation of organic MSW, and long-term creep. Coupled with underlying processes and heterogeneous waste properties, leachate injection operations—which are often employed to accelerate waste stabilization in landfills—can exacerbate differential settlements. Predicting these settlements is crucial to ensure the safety of landfills, while optimizing leachate recirculation operations is essential to prevent large differential settlements. The coupled-thermo-hydro-bio-mechanical (CTHBM) model, recently developed and validated, captures the key underlying mechanisms and provides a robust tool for reliable landfill performance assessment. However, this model becomes computationally expensive for large-scale simulations, limiting its applicability for uncertainty analyses and engineering optimization that require hundreds of simulations. To address this, the current study explores the application of machine learning (ML)-based surrogate models to predict surface settlements in MSW landfills. Four commonly used ML models were evaluated: random forest (RF), extreme gradient boosting (XGBoost), support vector regression (SVR), and multilayer perceptron (MLP). Surface settlement data from 273 landfill simulations with varying input parameters using the CTHBM model were collected. Data corresponding to eight spatial locations—on the surface of the landfill—was used for this analysis. The data was split in 80-20 ratio for training and testing the surrogate models. Bayesian optimization was used for tuning the hyperparameters to avoid potential overfitting. The predictive accuracy on test set, measured using the coefficient of determination (R²) values, for the resulting models was 0.96 for RF, 0.99 for XGBoost, 0.97 for SVR, and 0.99 for MLP. These results demonstrate excellent prediction accuracy, particularly for MLP—which showed good generalizability. Overall, this study demonstrates the effectiveness of ML-based surrogate models in accurately predicting landfill behavior while offering significant computational efficiency, thereby providing a robust framework for engineering optimization.