Data-driven prediction of landfill stabilization period using interpretable machine learning

Abstract

This study investigates the potential of data-driven machine learning (ML) models to capture the complex interdependent processes that control waste degradation in municipal solid waste (MSW) landfills and influence the time required for stabilization. Organic degradation in landfills can span several decades and is controlled by interdependent thermal, hydraulic, biochemical, and mechanical processes. While existing numerical modeling approaches can accurately simulate these processes, they are computationally intensive, limiting their feasibility for broader engineering optimization efforts. Moreover, data from laboratory or field studies, although valuable, are often site-specific and limited in scope, making them unsuitable for developing generalizable predictive models. To address these challenges, this study evaluates the feasibility of ML-based surrogate models trained on data generated from a coupled thermo-hydro-bio-mechanical (CTHBM) model developed at the University of Illinois Chicago. We evaluated several ML models − including support vector regression (SVR), tree-based ensemble models, and multi-layer perceptron (MLP) − for their accuracy and generalizability in predicting landfill stabilization periods. The tree-based ensemble models exhibited overfitting, with low accuracy on the testing data, indicating reduced generalizability. In contrast, SVR and MLP demonstrated high accuracy (R² ≈ 0.95), with root mean squared errors below one year on both training and testing datasets. Model interpretability tools, such as SHapley Additive exPlanation (SHAP) values and partial dependence plots, further illustrated the best-performing − SVR model’s efficacy in capturing the influence of key variables on the predicted landfill stabilization period. These results underscore the promise of ML-based approaches’ potential to enhance predictive tools for landfill evaluation and optimization.