Semiconductor manufacturing is one of the most complex, advanced, and competitive processes. The underlying factors include multiple product types, large production processes, re-entrant process flows, a large number of expensive and sophisticated machines, and batch processing.
To overcome the different challenges arising from the shop floor to the supply chain level, simulation-based approaches are widely used as they allow building a detailed model of the real system by incorporating realistic features. Recently, the idea of using a manufacturing simulation and supply chain environment has emerged, with the promise of offering the possibility to model, investigate and improve all relevant processes on a global scale in a risk-free world.
However, the quality of decisions based on simulations is determined by the quality of the model. Most manufacturing simulation models are designed for single-use applications and study a limited set of problems that are not reusable afterward. Another challenge that arises when developing a generic data-driven simulation is related to data management and information modeling aspects.
To obtain the necessary data for the simulation tool, required information such as bills of materials (BOM), resource capacities, process times, and other information, must be extracted from the enterprise requirements planning (ERP) system and manufacturing execution system (MES) with their database schema.
This paper proposes a generic, data-driven simulation model to evaluate and analyze a wide range of problems arising in modern semiconductor manufacturing systems.
Simulation Model Components
To effectively design and manage the complex and interdependent agents in the simulation model and to enable components reuse, the overall model is designed in a modular fashion with a separation of concern. This allows avoiding having an “all-knowing” agent responsible for most of the information about the simulation model.
Each agent addresses a separate concern regarding a high-level element of a semiconductor manufacturing plant. This approach gives more opportunities for agent upgrades, reuse, and independent development. Moreover, as the proposed model comes up with a visual animation, each agent can have a responsibility regarding elements rendering in the animation and a responsibility of managing business logic. Each agent is defined by specifying its business process as a discrete event model and interacts with other agents, thus generating the overall system behavior.
Data-driven Simulation Model Instantiation
A data-driven generic simulation model as “one which is designed to apply to a range of systems which have structural similarities” can simulate different instances of the systems in the domain without any change of the code. More precisely, a simulation model is automatically generated from an external data source using algorithms for creating the model and interfaces, which allows users to easily interact with the simulation environment without being aware of the code.
However, there is no uniform way to describe semiconductor plant data in meaning and context. It leads to non-standard naming conventions, vocabularies, and challenges regarding data interpretation. To fill this gap, researchers propose to move from proprietary fab-specific data description for common non-proprietary semantic descriptions by using ontologies formalized with the Web Ontology Language (OWL), a formal language for the description of terms and their relationship in a certain domain.
The «Main» agent queries the knowledge base and retrieves information about the different StationGroups presented in the plant. For each station group presented in the Knowledge Base, a new StationGroup Agent is instantiated. Then, each newly created StationGroup instance queries the knowledge base, retrieves the information about the StationsFamily that compose it, and instantiates them. Finally, in the same way, each StationsFamily queries the knowledge base retrieves the information of each station in this StationsFamily, and instantiates the Stations based on this information.
The initialization of the final step of the final route is the end of the construction of the data-driven simulation model.
2D simulation rendering
Simulation Model Dynamics
Besides the definition of the different model components, one of the biggest challenges of fab simulation modeling is related to modeling decision rules such as scheduling or dispatching rules, and other business rules. In the proposed model, two principal rules are defined:
find_StationFamily(): This function implements the scheduling policy by assigning a set of machines to a lot.
get_prioritary_lot_in_queue(): This function implements the dispatching policy. When several lots wait to be processed, this function is in charge of selecting the lot that will be processed.
These rules can be easily extended to implement and test different strategies.
Collecting data to populate a manufacturing simulation model can take a long time and can lead to several semantic integration problems, which can be difficult to detect. This problem can be solved using an ontology in the field of semiconductors.
For other research directions, researchers plan to use the proposed simulation model and semantic trends and initiatives in the semiconductor domain like SC3 to address in detail a wide range of problems arising in a semiconductor manufacturing plant and allow us to study their interactions.