Modeling a Resilient Supply Chain Using Machine Learning and Evolutionary Algorithms

Abstract

This study aimed to model a resilient supply chain using machine learning and evolutionary algorithms in the apparel, sporting goods, and electronic components industries. The research design employed a two-layer hybrid (predictive–prescriptive) framework with a quantitative, ex post facto approach. The population and sample comprised the standard DataCo database containing 180,519 logistics transactions, which was subsequently partitioned into three target subnetworks. Data analysis was based on the CRISP-DM methodology, K-means clustering, machine-learning models—particularly XGBoost (AUPRC = 0.8176)—and evolutionary multi-objective optimization algorithms (NSGA-II and MOPSO). NSGA-II demonstrated relative superiority in extracting a diverse Pareto front. The results showed that XGBoost provided acceptable performance in predicting delivery-delay risk, and the balanced strategy derived from NSGA-II yielded a proposed network design with an approximate cost of USD 327,650 and a mean risk of 0.519. Industry-level analysis further indicated that the sporting goods industry had the largest volume and cost, whereas the electronics industry exhibited the lowest cost and risk and therefore the most agile network. Future research is recommended to incorporate a disruption-scenario simulation layer to evaluate actual supply-chain resilience.