Simulation-Based Modular Scheduling System of Semiconductor Manufacturing

For example, Wein (1988) compared various release control strategies and scheduling rules and provided the better combinations of release control strategy and scheduling rule for three kinds of HP 24 models. Peyrol et al. (1993) implemented a simulated annealing methodology to determine the input order of a given set of products so as to minimize the average residence time of these products in the plant with unlimited intermediate storage policy. They developed a discrete simulation program to determine the value of the objective function. This method was applied to a semiconductor circuit fabrication plant of MOTOROLA Inc. Thompson (1996) described a simulation-based finite capacity planning and scheduling software to allow human planners to make better decisions. This tool improved cycle time, throughput, and equipment utilization significantly without adding additional equipment and personnel. Baudouin et al. (1995) presented information system architecture and a decision support tool to allow admissible scheduling solutions in a semiconductor device manufacturing fab.


Introduction
Discrete event simulation technology, imitating a real production environment by modelling, has been applied to the scheduling of semiconductor manufacturing fabrication facilities (fabs) since 1980s.
For example, Wein (1988) compared various release control strategies and scheduling rules and provided the better combinations of release control strategy and scheduling rule for three kinds of HP 24 models. Peyrol et al. (1993) implemented a simulated annealing methodology to determine the input order of a given set of products so as to minimize the average residence time of these products in the plant with unlimited intermediate storage policy. They developed a discrete simulation program to determine the value of the objective function. This method was applied to a semiconductor circuit fabrication plant of MOTOROLA Inc. Thompson (1996) described a simulation-based finite capacity planning and scheduling software to allow human planners to make better decisions. This tool improved cycle time, throughput, and equipment utilization significantly without adding additional equipment and personnel. Baudouin et al. (1995) presented information system architecture and a decision support tool to allow admissible scheduling solutions in a semiconductor device manufacturing fab.
Presently, discrete event simulation technology has been widely applied to predict the operational performance of a semiconductor manufacturing fabrication facility (fab) and assist its scheduling decisions. Existing fruitful research results can be classified into three main directions.
Simulation systems performing scheduling behaviors Some researchers developed various simulation systems to perform or assist scheduling behaviours of semiconductor wafer fabs. Ramírez-Hernández et al. (2010a, 2010b presented an architecture and implementation of a preventive maintenance optimization software tool, based on Approximate Dynamic Programming algorithms, for the optimal scheduling of preventive maintenance tasks in semiconductor manufacturing operations. Weigert et al. (2009) developed a simulation-based scheduling system for a semiconductor Backend facility. The heuristic search strategies were adopted to optimize the operating sequences with consideration on concurrent minimization of mean cycle time, maximization of throughput and due date compliance. Horn et al. (2006) developed a simulation-based interactive scheduling system, called "BackendPlanner", for the backend facility of a semiconductor manufacturer under practical conditions. Its special characteristics included the automated model generation from existent production databases, a heuristic optimization component, a fast-simulator, and the efficient data coupling strategy allowed the re-scheduling anytime, e.g., after machine breakdowns or other unpredictable events and several additional analysis tools reported important parameters of the manufacturing process, such as machine utilization, throughput time or work in process. Zhang et al. (2006Zhang et al. ( , 2009 implemented a simulation platform based on extended objectoriented petri net for real-time scheduling of semiconductor wafer fabrication systems, and designed a dynamic bottleneck dispatching algorithm to detect bottlenecks in a timely way to make adaptive dispatching decisions according to the real-time conditions. In view of semiconductor probe system's production scheduling problem, Zhang and Wang (2008) set up a probe behaviour model in terms of the figuration of UML and transformed it directly into simulation model by using the SimTalk of eM-plant. The experimental result of simulation model supplied the decision maker with evidence. Werner et al. (2006) developed a scheduling system for the backend of Infineon Technologies Dresden based on a discrete event simulation system and tested it in the real industrial environment. The simulation model was automatically generated from the databases of the manufacturer and was used for short term scheduling -from one shift up to one week. Sivakumar and Gupta (2006) conceptualized, designed, and developed a discrete event simulation based "online near-real-time" dynamic multi-objective scheduling system to achieve Pareto optimal solutions in a complex manufacturing environment of semiconductor back-end. The system used a linear weighted aggregation optimization approach for multiple objectives and auto simulation model generation for online simulation. In addition, it enabled managers and senior planners to carry out "what now" analysis to make effective current decisions and "what if" analysis to plan for the future.
Simulation as a tool to verify and improve the performance of scheduling algorithms Comparing to the achievements on the simulation-based scheduling system, the results of simulation as a tool to verify and improve the performance of scheduling algorithms are much more comprehensive. Jeng and Tsai (2010) applied simulation experiments to show their Match Due Date scheduling rule focused on the actual working mode of memory IC with the ability to reduce earliness and tardiness, make less number of setups and shorten the flow time and enhance the confirmed line item performance. Liu et al. (2010) presented dynamic scheduling models based on the resource conflict resolution strategies for a semiconductor production system and validated the model with a simulation case. Tang et al. (2010) proposed a genetic algorithm and simulated annealing algorithm based scheduling method in semiconductor manufacturing lines and verified it with a Minifab simulation model.

143
Chen and Wang (2009) proposed a modified fluctuation smoothing rule incorporating a fuzzyneural remaining cycle time estimator to improve scheduling performance in a semiconductor manufacturing factory, and evaluated its effectiveness by production simulation. Zhang et al. (2007) combined different dispatching rules and rework strategies to obtain better performance of cycle time, work in process, on-time delivery, and throughput. The simulation results showed the superiority of the proposed scheduling to static scheduling methods. Shi et al. (2008) used the simulation to determine the parameter values of a heuristic rule of a semiconductor manufacturing system and concluded a rule for the parameters' selection. Chou et al. (2008) developed a simulated annealing algorithm with a probability matrix integrated with a greedy heuristic to solve the dynamic scheduling problem of semiconductor burn-in operations optimally in practical sizes. They proved that the proposed method could effectively and efficiently obtain optimal solutions for small size of problems and provide high-quality solutions efficiently for large size of problems by simulating.

The evaluation and improvement of simulation-based scheduling methods
The importance of simulation-based scheduling methods has aroused many researchers' interests. They wanted to improve this method better than ever before.
For example, Li and Mason (2007) investigated the potential advantages and drawbacks of using simulation-based scheduling in a semiconductor wafer fab. Results suggested the potential for simulation-based scheduling approaches to improve on-time delivery performance to customers in wafer fabs. Koyuncu et al. (2007) augmented the validity of simulation models in the most economical way via incorporating dynamic data into the executing model, which then steered the measurement process for selective data update. Kim et al. (2003) proposed a simplification method for accelerating simulation-based realtime scheduling in a semiconductor wafer fabrication facility. In the suggested real-time scheduling method, lot scheduling rules and batch scheduling rules were selected from sets of candidate rules based on information obtained from discrete event simulation. Since a rule combination that gave the best performance may vary according to the states of the fab, they suggested three techniques for accelerating rule comparison, too.
Although the simulation-based scheduling approaches in semiconductor manufacturing field have made big progresses, their common problem is lack of a standard scheduling simulation model. Consequently, quite a few of research results are applicable to a special production environment and difficult to be directly extended to other semiconductor wafer fabs to obtain better performance. Presently, there are some exploratory researches on standard simulation model. For instance, Ralph et al. (2007) presented a prototype framework for scheduling of semiconductor manufacturing that divided the production system into Fabmodel, Process Route, Process Step, Tool Set and Operator Set. Rajesh and Sen (2004) proposed a method to simplify the simulation models' complexity.
In this paper, we design a standard simulation-based modular scheduling system (SMSS) by means of modularization and decoupling the algorithms from the models according to scheduling characteristics of semiconductor wafer fabs. The remainder of this book chapter is organized as follows. A general structure model (GSM) of the scheduling of semiconductor wafer fabs is introduced in section 2. Section 3 presents a data-based dynamic simulation modeling method. Section 4 designs and develops a Simulation-based Modular Scheduling System (SMSS). Finally, we validate SMSS with a case study in section 5. Section 6 contains conclusions and future works.

General structure model
In view of its physical structure, a semiconductor wafer fab is composed of dozens of workcentres. Each work-centre includes multiple machines. Every machine has several or dozens of recipes to finish more than one operation of a job. An operation of a job can be finished by one or more machines and its processing time on different machines may be different.
In view of its process flow, a job will be processed as soon as it is released to a wafer fab. The machine processes a job according to an existing sequencing plan or directly selects a job in its queue according to some priority rule considering the real states of the wafer fab. A job's processing record is stored in the manufacturing execution system (MES). A machine needs a preventive maintenance at regular intervals.
In view of product definition, a job has its own process flow in a wafer fab. There are hundreds or thousands of jobs in a wafer fab at the same time. They may belong to different product editions. If the process flow of each product edition is stored in MES, the data storage will be huge. And the process flow of a product edition is difficult to be modified. In addition, a process flow is composed of hundreds of operation sets. An operation sets includes several operations. An operation is corresponding to several pairs of machine and recipe. An operation of one job may be the same with that of other jobs. So we only need storing these operations in MES and combining them into different process flows.
In view of scheduling of a semiconductor wafer fab, its objective is to reasonably utilize the machines and resources in the fab to achieve better operational performance and meet the requirements of the jobs' process flows and other constraints as well. In other word, the scheduling behaviour is responsible for arranging right resources for an operation at right time. There are three main scheduling ways in a semiconductor wafer fab, i.e., static scheduling (usually called sequencing), dynamic dispatching and rescheduling. Static scheduling is to determine the processing order and time for specified operations before the real processing begins. Dynamic scheduling is to determine next processing route of a job or a processing priority of a job on a machine according to current states of a fab in a real-time style. Due to its uncertain manufacturing environment, an existing sequencing plan for a wafer fab may be not applicable to implement. Then the sequencing plan must be modified or regulated to be suitable to the real production environment. This behaviour is called rescheduling.
Based on the analysis results as above, the scheduling of a semiconductor wafer fab can be abstracted as four parts, i.e., WIP-centered configurable definition, machine-concerned physical environment, process information recording production information, and scheduling describing detail algorithms. Then the general structure model of scheduling of a semiconductor wafer fab can be divided into static part and dynamic part (shown in Fig.1).
The static part defines static information of a semiconductor wafer fab, including physical (machine-centred) and configurable (product-centred) definition. The dynamic part defines the information used during the production, including process information (e.g., release plan and sequencing plan) and scheduling related information (e.g., scheduling algorithms). Configuration definition regulates the production on the physical machines by defining the process of WIP. It also offers reference to the scheduling algorithms. Process information provides information (usually a sequencing plan) to machines to facilitate their production processes. On the contrary, the processing on a machine generates process information, too. In addition, process information acts as the input or feedback of scheduling algorithms. Transferring those parts to its concept layer, a GSM of scheduling of semiconductor wafer fabs is also defined as four layers, i.e., configurable definition layer, physical layer, process information layer and scheduling layer. The components in each layer and the relations between these layers are shown in Fig.2  The configurable definition layer manages process information of the products of a semiconductor wafer fab, such as products, processes, operation sets and operations. The minimum unit is the operation. Multi-operations compose an operation set, and multioperation sets constitute a process. Every product has its own process specifications.
The physical layer records resource information of a semiconductor wafer fab, such as its layout and machines, the process abilities and state information of these machines, and state information of WIPs.
The process information layer records the data generated from the simulation processes, including input and output information. Input information contains release plans and machine maintenance plans. Output information includes dispatching lists, machine maintenance tasks, performance issues and the work log of the machines.
The scheduling layer contains some algorithms used by the simulation, including release control algorithms, scheduling algorithms and rescheduling algorithms. Release control algorithms generate the release plan during a special period. Scheduling algorithms, applied to the simulation running, generate the dispatching lists and the maintenance tasks for the machines during a special period. Rescheduling algorithms are used to modify the existing dispatching lists or generate new dispatching lists for a special period.

Data-based dynamic simulation modelling method
There are many distinctions between different semiconductor wafer fabs. As a result, it is very difficult to abstract their commonness. The traditional simulation modeling method, focusing on a special semiconductor wafer fab, has strong pertinence but less generalness. Correspondingly, the release control and scheduling algorithms used in a traditional simulation model only fit to a special simulation model, which are difficult to be applied to other simulation models to obtain better performance. In addition, it is not easy to evaluate those algorithms' performance under different production environments.
The less generalness of the traditional simulation models can be sol ved by using the proposed GSM in section 1. As for the strong pertinence, it is noticed that the differences between the semiconductor wafer fabs are their layouts, machines, processes and scheduling methods that can be transferred to data by some ways. If it is built dynamically by using these data, the simulation model is merely dependent on these data, which overcomes the deficiency of the strong pertinence.
Data types for dynamic modelling The process of dynamic modelling of the scheduling simulation model of semiconductor wafer fabs is to upload data, handle data, and finally organize data into a simulation model with specified structure. Each data unit belongs to some module in a layer of GSM. So the first job to analyse the require data types for dynamic modelling of a simulation-based scheduling model is to analyse the data requirements of each module in GSM.
Based on the definition on GSM, data types required by the simulation-based scheduling model are summarized into Product data is corresponding to the configurable definition layer in GSM. Because every wafer fab has its own products' definition, the product data is required without doubt for dynamic simulation modelling. Product data includes information related to products and relations between operation sets and machines. The latter defines the relations between physical layer and configurable layer in GSM. •

Physical resource data
Physical resource data is corresponding to the physical layer in GSM. It describes the physical layout of a semiconductor wafer fab. Because different wafer fabs have their own physical layout, the physical resource data is required without doubt for dynamic simulation modelling, too. Physical resource data is the main body of dynamic simulation model. It decides the numerical layout of a wafer fab in computer software. Traditional modelling methods directly set physical resource data into a special simulation model. If there are some changes (e.g., new machines addition) in the layout, the simulation model must be modified. However, data-based modelling is to build the numerical layout of a wafer fab dynamically according to physical resource data. Therefore, if there are changes in the physical layout of a wafer fab, no extra work is required during modelling process. The dynamic data-based modelling method is more flexible and general than traditional modelling methods. •

Process data
Process data, corresponding to the process information layer in GSM, includes the information required by production processes, e.g., maintenance tasks and release plans. It is the precondition to run the simulation. Process data doesn't include information about the historical records, sequencing plans and performance records that are generated during the simulation running process. So these data is unrequired during model uploading process. •

Algorithm data
Algorithm data, corresponding to the scheduling layer in GSM, records the scheduling algorithms applied to a wafer fab. The detailed scheduling algorithms are realized in the simulation model. Algorithm data records the information related to the algorithms, including algorithm names and parameters.
Once these four kinds of data are self-contained, the simulation model of a semiconductor wafer fab can be loaded dynamically successfully. The flexible superiorities of dynamic data-based simulation model comparing to traditional simulation model are as follows.
Firstly, these four data types are relatively independent and can be easily combined to build simulation models of different semiconductor wafer fabs. So the reusability of these data types is enhanced and the workload of simulation system development is seriously reduced.
Secondly, it is much easier to modify a simulation model. To modify a simulation model needs only the modification of the data required by generating the dynamic simulation model.

Dynamic modelling process
In view of traditional simulation modelling method, its structure (i.e., physical layer) is determined at the beginning of building the simulation model. Data is added to the physical layer. So if there are some changes in the physical layer, the data and model need corresponding modification with heavy workload. This kind of modelling process is called static modelling (shown in Fig.3). The data of static modelling will not be uploaded unless the simulation process starts. In view of dynamic data-based simulation modelling method, data takes an important role during the modelling process. On the contrary, the simulation model is an assistant. All of the changes of the simulation model are realized by changing the data. In other word, the simulation model of the same semiconductor wafer fab may be different due to the changes of its data. This kind of modelling process dependent on the data is called dynamic modelling (shown in Fig.4). The data-based dynamic simulation modeling flow is as follows: load the original data first, then deal with these data to standardize them into structural data required by dynamic modelling, and finally organize these structural data into a simulation model with a specified structure. These data will attribute to some modules of some layers of GSM ultimately.

Simulation-based modular scheduling system
Based on the proposed dynamic simulation modeling method in section 3, we design and develop a simulation-based modular scheduling system (SMSS). SMSS integrates six semiconductor wafer fab models, including 4-inch wafer fab (Bl4), 6-inch wafer fab (Bl6), HP24Fab1, HP24Fab2, HP24Fab3 and Mini-FAB. Modular means that SMSS integrates various kinds of scheduling in a wafer fab, such as release control, sequencing, dynamic dispatching and rescheduling methods. Each scheduling method is encapsulated as a module. The modules belonging to the same kind of scheduling can be replaced each other.
The hardware environment of SMSS is Pentium IV 1.5 CPU, 256M memory and 10G disk driver or more. The software environment includes Windows 2000 or XP operation system, .NET Framework 2.0, Microsoft Office 2000 or higher edition, Graphics.Server.NET.v1.1 and simulation platform eM-Plant.
The architecture of SMSS is shown in Fig.5. There are three layers in SMSS, i.e., data layer, software layer and simulation layer.  There are communications between data layer and software layer. There are communications between data layer and simulation layer, too. These communications are enabled through corresponding data interfaces to facilitate free transfer between different semiconductor wafer fab models.
The design principle of data layer is to guarantee identical data structures of six semiconductor wafer fab models. The standard of the identity is the data types in Table 1. To meet the requirements of GSM on data, we build following data-entity tables in the ACCESS database.
• The relations between above-mentioned data-entity tables and data types in Table 1 are shown in Table 2.

Software layer
Software layer is the core of SMSS. It is used to model a semiconductor wafer fab in a standard way, manage SMSS's modules configuration, and provide a user management interface. Through specified data interface, software layer can read data from data layer.
Then it configures the modules in it and sets the parameters of modules. Finally, it can call simulation layer to realize the simulating process of semiconductor wafer fab models.

Data types Data-entity tables Product data
Product table  Physical resource data  Work-centre & machine table, WIP table   Production data  Release table, maintenance task table, dispatching list  table, performance table  Algorithm data  Scheduling algorithm table   Table 2. Relations between data-entity tables and data types.

Simulation-Based Modular Scheduling System of Semiconductor Manufacturing 151
Software layer includes five components, i.e., data interface definition, general structure model, scheduling modules, simulation interface and user interface. In addition, we add some functions to make statistics on system data, generate their diagrams and files, and manage the system parameters. •

Data interface definition
Data interface definition is to read data from data layer to software layer and organize these data according to the data types in GSM (shown in Fig.6). Data layer stores data from six semiconductor wafer fab models. Although their data are different, the structures of their databases are the same.
So we can only develop one data interface to read either model. If there are other databases with different structures, we only need to define corresponding data interfaces to integrate these heterogeneous databases easily.
• General structure model General structure model organizes the data read from data layer into data types described in Table 1 to facilitate scheduling modules to call them quickly.  •

Scheduling modules
The scheduling algorithms in MPSS are realized as multiple modules. A scheduling module obtains the algorithm configuration information from user interface, then configures the algorithm's parameter and implements the computation process, finally sends the computation results to the user through the user interface. •

Simulation interface
Simulation interface is the communication interface between software layer and simulation layer. Its functions include: the open, close and save of simulation models, simulation control (i.e., simulation start, pause and reset), and implementation of simulation inner script language. •

User interface
User interface is the interactive interface for user operating software layer. The design principle of use interface is full-open, i.e., the panels for setting parameters are all on the main frame and reducing the number of menus as less as possible.

Simulation layer
The simulation layer is to simulate a current semiconductor wafer fab model. It takes eM-Plant 7.5 as its platform, including a model base, an algorithm base and a data interface. •

Model base
There are six semiconductor wafer fab simulation models in the model base. These models are generated dynamically according to the method introduced in section 3. So the structure, control programs and data interface of these models are identical. •

Algorithm base
There are four parts in algorithm base: model uploading algorithm, simulation model control algorithm, scheduling rules and statistical algorithm. These algorithms are applicable to all simulation models in the model base. •

Data interface
Data interface is used to interact with database to read original data from data layer and write simulation data back to the database.

Case study
SMSS can be used as a platform to test the performance of different scheduling algorithms and the proper mix level of scheduling algorithms. Thus, it can be considered as a reference tool for the production decisions.
The workflow of SMSS includes uploading simulation model, generating release plan, generating sequencing plan, dispatching and rescheduling (shown in Fig.7).

Uploading simulation model
The uploading simulation model process is to transfer data from data layer to software layer. Due to huge amount data in semiconductor manufacturing, this process may spend longer time. The uploading time for six simulation models is shown in Table 3.

Release control modules
There are four release control modules, i.e., ConNum (the number released to the fab per day is the same), HybridIntelligent (the release plan is generated by an immune algorithm), PredictDueDate (the jobs released to the fab according to the urgent level of their expected due dates) and MultiObjective. Taking Minifab as an example, the performance of each release control modules are shown in Table 4.   Table 4. Performance statistics of release control modules.

Scheduling modules
The scheduling algorithms in SMSS include FIFO, EDD, EODD, LPT, SPT, CR, FSVCT, LS, FIFO+, SRPT, SRPT+, SRPT++, FSVL and DBR. The release period is set to 4 months. We select the data in the middle 2 months to implement analysis work. The judge criterion is set to the product of the on-time delivery and the throughput. The former 10 matching of the release control algorithms and the scheduling algorithms are shown in Table 5.
Obviously, the matching of HybridIntelligent and DBR is the best decision. It can be seen from the simulation results that SMSS has a strong ability to support impersonal evaluation on different scheduling algorithms and release control modules. It will take an important role on assisting the operation management of complex production environments.

Release control algorithm
Scheduling algorithm

Conclusion
This book chapter introduces a simulation-based modular planning and scheduling system of semiconductor manufacturing (SMSS). The main advantages of SMSS are as follows.
1. It adopts a data-based dynamic simulation modelling method to be adaptive to various production environments. 2. It applies a module-based decouple of the algorithms from the simulation models to enhance the simulations' efficiency and make simulation models' to be reused. 3. It offers a better impersonal evaluation platform for further research on the scheduling algorithms.

Acknowledgment
This work was supported in part by National Natural Science Foundation of China (No.50905129, No. 61034004) and Program for Scientific Innovation of Shanghai (No.09DZ1120600, 10DZ1120100). The authors would like to thank the editors and anonymous reviewers for their valuable comments.

References
Baudouin, M.; Ruberti, C.; Arekion, J. & Kieffer, J. P. (1995). Decision Support System Based on a Factory Wide Information Integrated System and Discrete Event Simulation to