توزیع زمان واقعی چند هدفه برای تحویل یکپارچه در یک فاب با استفاده از بهینه سازی شبیه سازی مبتنی بر الگوریتم ژنتیک

In a wafer fabrication Fab, the “integrated delivery”, which integrates the automated material handling system (AMHS) with processing tools to automate the material flow, is difficult to implement due to the system complexity and uncertainty. The previous dispatching studies in semiconductor manufacturing have mainly focused on the tool dispatching. Few studies have been done for analyzing combinatorial dispatching rules including lot dispatching, batch dispatching and automated guided vehicle (AGV) dispatching. To handle this problem, a GA (genetic algorithm) based simulation optimization methodology, which consists of the on-line scheduler and the off-line scheduler, is presented in this paper. The on-line scheduler is used to monitor and implement optimal combinatorial dispatching rules to the semiconductor wafer fabrication system. The off-line scheduler is employed to search for optimal combinatorial dispatching rules. In this study, the response surface methodology is adopted to optimize the GA parameters. Finally, an experimental bay of wafer fabrication Fab is constructed and numerical experiments show that the proposed approach can significantly improve the performance of the “integrated delivery system” compared with the traditional single dispatching rule approach.

Semiconductor manufacturing is one of the most sophisticated manufacturing processes with characteristics of large processing steps, re-entrant material flows, and batching processing. In the past decades, the dispatching decision problems of processing tools and material handling in wafer fabrication are usually investigated separately due to the modeling complexity. However, an effective dispatching mechanism which properly integrates tools dispatching with AMHS dispatching is becoming a more widely used strategy to achieve high performances in semiconductor wafer fabrication systems. Most approaches to the semiconductor manufacturing scheduling problem can be classified into four categories: heuristic rules, mathematical programming techniques (such as branch and bound, Lagrangian relaxation and queuing network model), neighborhood search methods (such as Tabu search, genetic algorithm and filtered beam search) and artificial intelligence techniques (such as artificial neural networks and expert/knowledge-based systems) [1]. In recent years, many studies have focused on the dispatching rules for scheduling semiconductor wafer fabrication systems. Heuristic rules and mathematical programming techniques are widely used for this problem. Yang and Chang formulate a multi-objective model for IC (Integrated Circuit) sort and test [2]. In their study, Lagrangian relaxation is used to look for an approximate Pareto boundary and a new algorithm is designed to solve the dual problem. Li et al. propose a dispatching rule to improve on-time delivery without decreasing throughput and increasing cycle time [3]. Wu et al. develop a line balance-starvation avoidance (LB-SA) algorithm based on a proposed simplification model of the processing route for a Fab with machine-dedication features [4]. Monch and Driebel also propose a heuristic rule, named modified shifting bottleneck heuristic, for wafer fabrication Fab [5]. Among these approaches, dispatching rules, by far, are the most commonly used tools for shop floor control. The reason behind this is that they are simple to implement, quick in reacting to the changes encountered on the shop floor, easy to understand and require a low computational load [6]. Pierce and Yurtsever present a value-based concept. They mainly concentrate on generating profits [7]. Based on that, Hsieh and Hou develop a production-flow-value-based job dispatching rule (PFV) by the theory of constraints (TOC) for wafer fabrication [8]. The TOC and profitability costs estimation of a WIP-wafer lot are derived to prioritized jobs based on their profitability. Although dispatching rules are usually sub-optimal and myopic, over the years, researchers have successfully introduced advanced rules that are capable of improving multiple performance measures simultaneously [6]. Dabbas and Fowler develop a composite dispatching rule for each station, which combines multiple dispatching rules, including local and global rules, into a single rule [9]. It is designed to simultaneously maximize the on-time delivery and minimize the variance of lateness and cycle time by using a linear combination with relative weights. Bahaji and Kuhl also propose a composite dispatching rule and prove it to be robust for the average and variance of flow time, as well as due-date adherence measures [10]. Another new weighted composite rule suggested by Yang et al. also take into consideration the general advantage of the earliest due date (EDD) and shortest processing time (SPT) rules, with the weights determined from the historical data [11]. However, those rules above are designed for lots in single processing stations and may not perform the best for different scenarios. Overhead monorail systems are widely used in wafer manufacturing systems due to their efficiency in lean designs, gentle handling of wafers and efficient use of overhead space [12]. It has been demonstrated that they can significantly improve the inventory storage and material handling system reliability [13]. Christopher et al. indicate that both processing tool and vehicle dispatching rules and their interaction have a significant impact on Fab performance if the AMHS is not extremely over or under utilized. It also has been shown that the combination of dispatching rules is highly dependent on the specific Fab [14]. In order to solve the scheduling problem that integrates processing tools and vehicles, Lin et al. propose a hybrid push/pull dispatching rule, including the procedures for vehicle selecting lot (VSL) and lot selecting vehicle (LSV), to improve the photobay performance [15]. Min and Yih propose a combined simulation and neural network approach in which simulation is used to collect data related with the system status, performance and change in dispatching rules, and then its results are fed into a neural network to obtain related knowledge [16]. Tyan et al. propose a state-dependent policy for achieving better system performance [17]. However, these methods are not capable of handling complex stochastic semiconductor manufacturing environments. Through a survey on the use of discrete event simulation for manufacturing system design and operation, Smith shows that simulation has proven to be an extremely useful analysis tool for developing dispatching rules [18]. Um et al. gives a simulation design and analysis of a flexible manufacturing system with an automated guided vehicle system (AGVs) [19]. Hung and Chen explore a simulation-based dispatching rule and a queue prediction dispatching rule for searching better dispatching rules to reduce flow times, while maintaining a high machine utilization [20]. Kim and Jeong present a simulation-based real-time scheduling methodology which considers both simulation models and decision time points for reselecting new rules [21]. Sivakumar also develops a discrete event simulation-based on-line multi-objective scheduling approach, which includes the use of a linear optimization algorithm and automatic simulation model generation, in a complex manufacturing environment [22]. Kim et al. present a simulation-based real-time scheduling (SBRTS) methodology in which lot scheduling rules and batch scheduling rules are selected from candidate rules based on information obtained from the simulation [23]. Jelong et al. present a hybrid approach that combined GA with the simulation [24]. However, only single performance indicator, namely the maximum completion time for the last job, is optimized in their study. The contribution of this paper is twofold. First, it mainly concentrates on the “combinatorial dispatching” which has rarely been studied. Different from previous study on single performance indicator or weighted composite dispatching rules, a representative Pareto optimal solution subset is explored to handle multiple objectives. A GA-simulation procedure is applied to obtain this solution subset. This methodology could provide users the opportunities to select appropriate solutions according to their preferences. Second, a GA parameters optimization process is developed to handle complex environments. This process can improve the performance results by optimizing GA parameters even with the dramatically changing environments. The remainder of this paper is organized as follows. In Section 2, the proposed methodology is given. Related dispatching rules, GA parameters optimization by response surface methodology and simulator parameters optimization will be presented in this section. Sections 3 and 4 describe GA controller and the simulation model in details, respectively. In Section 5, numerical experiments used to show that the proposed methodology can efficiently improve the overall system performance. Finally, Section 6 concludes the paper.

This paper proposes a GA based simulation optimization methodology to solve a real-time multi-objective dispatching decision problem for the “integrated delivery” in a 300-mm semiconductor wafer fabrication Fab. The methodology, consisting of the on-line scheduler and the off-line scheduler, can optimize the combinatorial dispatching rules including lot dispatching, batch dispatching and AGV dispatching. Four modules (GA parameters optimization module, simulation parameters optimization module, GA controller and simulation model) compose of the off-line scheduler. They are used to optimize the parameters for searching better combinatorial dispatching rules. A numerical experiment is designed to verify if the GA parameters optimization module is able to improve the performances and another experiment is designed to compare the proposed methodology with the widely used single dispatching rule and weighted sum method. It has been seen that the proposed methodology outperforms single dispatching rule in overall performance indicators and can save much more time compared with weighted sum method. In order to deal with the multi-objective problems, a representative Pareto optimal solution subset is required. Since it may be hard for users to determine a solution from the subset, a good mechanism for finding a satisfied subset or selecting appropriate solutions for users will be the future research direction.