摘要
Overall architecture of Deep reinforcement learning (DRL) based Digital Twin (DT) for production planning.
[Display omitted]
•A deep reinforcement learning based Digital Twin framework is introduced for dynamic production planning strategies.•The proposed approach can reduce production risks, costs and enhance demand fulfilment rate.•The proposed approach is validated with realistic data in different demand scenarios.•The developed solution has been implemented in a semiconductor company.
Increasing needs for demand fulfilment in semiconductor manufacturing while facing shortening product life cycles, high product mix and complex capacity portfolio are challenging. The importance of production planning increases to meet the goals of timely delivery and lower production costs. Traditional production planning systems may not be sufficient to solve challenges including adaptive issues while facing high product-mix, complex capacity portfolio and the rising demand fluctuation in semiconductor manufacturing. Limitations of existing approaches can be traced in part to the lack of a framework within which fast responsive and adaptive system interaction can be integrated in light of demand uncertainty and production dynamics. This study aims to propose a deep reinforcement learning based Digital Twin framework to provide fast response, adaptive virtual reality interaction and dynamic lot release strategies while reducing production costs and enhancing demand fulfilment rate from production dynamics. To provide fast responsive and adaptive system, the approach collects information from manufacturing environment for the twin simulation model for deep reinforcement learning to select the most appropriate production planning models while considering operational demand uncertainty. To meet realistic needs, an empirical study in semiconductor wafer fabrication is conducted. The performance of the proposed approach outperforms other stochastic models. The proposed approach contributes to practical viability to for self-adaptive system and provides more resilient production plans under production dynamics.