An explainable framework integrating machine learning and life cycle assessment for balancing performance, carbon emissions, and costs in asphalt mixtures

Abstract

This study proposed an optimization method for Non-Dominated Sorting Genetic Algorithm II (NSGA-II) and applied the optimized NSGA-II to the multi-objective optimization problem involving the performance, carbon emissions, and costs of asphalt mixtures. Firstly, the Marshall stability (MS) predictive equation of asphalt mixtures was established. By leveraging life cycle assessment (LCA), the asphalt mixtures were partitioned into three stages: raw materials production, transportation, and mixtures mixing. The carbon emissions and costs calculation objective functions for asphalt mixtures have then been constructed. Subsequently, a hybrid machine learning model combining the NSGA-II and Cauchy distribution crossover operator (NSCCGA-II) was developed to address the issue of easily getting trapped in local optima that exists in NSGA-II. Through multiple evaluation indicators, it is demonstrated that the solving performance of NSCCGA-II is superior to that of NSGA-II. The solutions of the balanced design model were obtained and evaluated. The findings indicated that the MS of the acquired Pareto solution set are all above 11 kN, the carbon emissions are all below 57 kg, and the costs are all below 595RMB. When the weights of the three optimization goals are set to be equally important, the basalt fiber (BF) content, nano-TiO<inf>2</inf>/CaCO<inf>3</inf> (NTC) content, and asphalt-aggregate ratio of the top-ranked Pareto solution are 3.93%, 5.57%, and 4.35%, respectively. The corresponding MS, carbon emissions, and costs are 12.22 kN, 53.62 kg, and 558.80RMB, respectively. Finally, the importance analysis of Gradient Boosting Regression (GBR) and Shapley Additive Explanation (SHAP) indicated that asphalt-aggregate ratio is of the highest importance to MS and is much higher than BF content and NFT content.