Digital twin model with machine learning and optimization for resilient production–distribution systems under disruptions

Abstract

Inspired by a real-life problem in the semiconductor industry, we introduce a novel digital twin model for a company subject to the adverse effects of unpredictable disruptions. Specifically, this company manufactures a product using a raw material provided by an external supplier, whose lead times may abruptly change due to disruptive events. The Smoothing Order-Up-To rule is adopted by the company as a replenishment policy. It is characterized by three control parameters, which must be optimized to enhance the resilience of the system. To this end, the digital twin learns from the real production–distribution data and periodically self-adjusts the replenishment parameters based on the evolution of the external environment. The digital twin architecture combines data analytics, simulation modeling, machine learning, and a metaheuristic. More specifically, an Artificial Neural Network learns from the manufacturer’s operations and generates predictive models. These are embedded in a Particle Swarm Optimization, which provides the optimal combination of the replenishment parameters. An experimental campaign was performed to demonstrate that the digital twin outperforms the traditional strategy in which the replenishment parameters are kept unchanged. The numerical results show that the digital twin strongly improves the manufacturer’s performance, in particular in terms of time-to-recover and time-to-survive, used to measure the resilience of the system subject to disruption.