Predictive stochastic repurposable supply chain: an adaptive strategy toward supply chain plasticity

The supply chain resilience concept traditionally implies disruptions that occur on a predictable scale and with specific durations, and assumes recovery activities will begin when the disruption ends. However, long-term crises such as trade wars and pandemics are generally unpredictable, persistent over long periods, and require that their recovery activities be begun during the crisis itself. In this case, returning to the old normal is often unrealistic and infeasible; thus, the supply chain must adapt to the new normal to maintain survivability and long-term viability. The present study introduces a new concept in supply chain management, the Repurposable Supply Chain (RSC), in which the supply chain adapts to long-term crises by shifting its operations to emerging product networks. In this regard, a two-stage stochastic MILP model is developed under uncertain disruption severity over the planning horizon. Since scenario generation is a complex process in such long-term crises, a predictive algorithm, equipped with machine-learning models that use real-time crisis-severity data, prunes the scenario tree and estimates the probabilities of the remaining scenarios. To solve the proposed RSC stochastic optimization problem at a large scale, a Generalized Benders Decomposition approach is developed and implemented. By implementing the proposed predictive algorithm, the number of potential scenarios is reduced from 729 to 8, which shows substantial pruning of the scenario tree. Income analysis on the case study reveals that the RSC performed at a higher level of the non-repurposable model, resulting in a 9.7% increase in total income, which relevant sensitive analysis proves the robustness of the result. Purchasing cost has been identified as the most sensitive term, with a 20% increase raising the objective value by 7.5% in the RSC compared to 8.8% in the non-RSC in the proposed case study. Stochastic analysis reveals that the learning-powered RSC more effectively handles uncertainty and captures its impacts. Finally, the proposed solution method effectively handles large-scale RSC instances within reasonable time limits and shows significant computational performance.