An empirical hybrid framework combining system dynamics, structural equation modeling and machine learning for circular economy transitions in urban solid waste management in Vietnam

Transitioning to a circular economy in urban solid waste management requires simultaneous changes in household source-separation behavior and in the technical infrastructure for collection, recycling, and treatment. This study develops and empirically evaluates a three-layer hybrid framework combining structural equation modeling, system dynamics, and machine learning. The behavioral layer extends the theory of planned behavior and is estimated using primary household survey data collected in two waves from the same households in first-class Vietnamese cities. The longitudinal design enables the study to estimate the lagged relationship between observed cumulative recycling performance and subsequent perceived collective efficacy, while the contemporaneous behavioral relationships are estimated using partial least squares structural equation modeling. The material-technical layer is a system-dynamics model calibrated using official operational data on waste generation, collection, recycling, and landfilling. The empirically estimated behavioral and longitudinal relationships are then embedded in a coupled model in which recycling outcomes feed back into subsequent perceived collective efficacy. The machine-learning layer combines gradient boosting regression and an artificial neural network trained on Monte Carlo simulations of the coupled model and provides a computationally efficient surrogate for global sensitivity analysis and multi-objective policy optimization. Applied to first-class Vietnamese cities over the 2025–2040 horizon, the bidirectional-feedback specification reaches a cumulative recycling rate of 74% by 2040, approximately 12–15% points higher than the corresponding one-way coupling, and reaches the 70% recycling threshold two to three years earlier. Global sensitivity analysis identifies infrastructure subsidy level and social-norm communication intensity as the two most influential policy parameters, with total-order Sobol indices of 0.31 and 0.24, respectively. Multi-objective optimization identifies a policy combination that reaches the same recycling target at approximately 18% lower discounted cost than a strategy relying on infrastructure investment alone. The study contributes an empirically grounded hybrid architecture in which the behavioral relationships and the temporal feedback mechanism are estimated from observed household and operational data, while the 2025–2040 policy outcomes are generated through scenario-based simulation.