| Summary: | In semiconductor manufacturing, fast and effective dispatching is important. Effective dispatching improves factory performance by dispatching wafer lots that accommodate production priorities. Dispatch rules are used to schedule lots in the factory. However, manufacturing situations are not static but dynamic, with frequent shifts in the product mix and ad-hoc events such as equipment breakdown. Earlier studies have shown that changing dispatch rules by reacting to a dynamic manufacturing situation improves the overall factory performance. Also, it has been established that no one dispatch rule consistently outperforms other dispatch rules for all manufacturing situations. This underscores the need for a dispatching system that adapts to manufacturing situations quickly, and makes effective decisions.
This thesis introduces a machine learning-based, situation-aware dispatching system to adapt dispatch rules in dynamic manufacturing situations. The system has two key components; i) Best Rule Learner, that identifies the best dispatch rule for a manufacturing situation; ii) Factory Situation Generator, that generates a wide array of synthetic manufacturing situations for increased robustness of the Best Rule Learner.
The Best Rule Learner defines and identifies features that symbolize the manufacturing situation. The features are generated by snapshot data, which represents the manufacturing situation at a specific time; and time series data, which represents changes over periods. Snapshot data provides a static view of the manufacturing situation at a specific moment, while time series data captures dynamic changes over periods, allowing for a comprehensive representation of the manufacturing situation when combined. A multi-pass (discrete event) simulation technique is used to generate training data corresponding to the chosen features. This technique involves running simulations to collect the performance impact of different dispatch rules at the same decision points as the training dataset. Machine learning model is then used to learn the relationship of the manufacturing situation and the best dispatch rule from this dataset. A series of experiments are conducted with a scaled down semiconductor factory model to determine features that could accurately represent manufacturing situations which lead to a gain in factory performance. The experiments conclude that combining snapshot data and time series data increases the effectiveness of the Best Rule Learner.
The Factory Situation Generator, is used to generate a variety of manufacturing situations in order to improve the robustness of the Best Rule Learner. Different approaches from other domains are adapted and a series of experiments are conducted to identify the best approach that generate diverse and distinctive synthetic manufacturing situations semiconductor manufacturing. The experiments conclude that Generative Adversarial Networks (GAN) is able to meet this objective.
The situation-aware dispatching system presented in this thesis is important to eliminate human-in-the-loop in adapting dispatching rules with changing manufacturing situations. With increasing complexity in semiconductor manufacturing due to cutting edge products and volatile demand, such a system is the need of the hour.
|
|---|