晶圓代工廠考慮週期時間限制之機台規劃研究

Abstract

機台規劃問題是決定一個工廠需採購的機器種類與數量，以達成企業所設定的目標。一個好的機台規劃方式可以決定較佳的機台採購計畫，進而降低機台採購成本並提高生產績效。因為晶圓代工業者在規劃機台時，所考慮的問題面並不相同，因此我們機台規劃的內容乃分成三個主題來探討。第一個主題是針對不確定需求的環境來探討機台規劃方式；第二個主題是針對有季節性產品調整的業者，建構一個多產品比例的機台規劃方式。第三個主題是分析擁有多座舊廠的業者，其建新廠時的機台規劃問題。 近年來有許多的研究著手探討不確定需求下的機台規劃主題，然而這些研究大多未考慮生產的週期時間。因為週期時間是晶圓廠營運績效的重要指標，我們認為在規劃機台組合時它是不可或缺的考慮因素。所以本論文第一個主題，便是探討不確定需求下有週期時間限制的機台組合決策。模型中我們使用一組需求情境來代表未來不確定的需求，而每一個需求情境均對應一機率值，代表此產品需求未來可能出現的機率。在此不確定需求情況下，生產的週期時間不超過一既定目標週期時間，我們決定一組最適機台組合以達成利潤最大的目標。 第二個主題的研究動機乃是為了解決某些業者產品有季節性調整的問題。許多晶圓代工廠的訂單常常會隨著季節的不同而產生變化，因此造成晶圓代工廠在一年當中需要生產多種產品比例，所以工廠的機台需能夠配合生產不同的產品比例。此主題的機台規劃模型，也考慮了週期時間與機台採購預算的限制，目標是決定一組最適機台組合使得營業利潤最大。 第三主題是針對一個晶圓業者目前已擁有多個晶圓廠，然而計畫增建一座新廠的機台規劃問題。雖然舊廠目前均有自己的特定產品產出比例，但為了使新廠的機台成本最低，我們在規劃新廠機台時須重新決定每一個分廠的產品產出比例。本主題的決策變數包括新廠的機台數量及各分廠的產量配置，限制式包了週期時間因素，目標是使新廠的機台投資成本最低。 我們以基因演算法來求解上述的問題，並使用等候模型來評估機台的績效。實驗結果顯示，本研究的所求得的解較其他規劃方式為佳。

The tool planning problem is to determine how many tools should be allocated to each tool group to meet some objectives. An effective planning will provide a toolset, which has lower cost for procurement and yields good performance. This paper addresses three tool planning problems faced by semiconductor manufacturers. The first topic describes a tool planning problem under uncertainty in demand. The second topic presents a tool planning in multiple product mix scenarios. The third topic proposes a tool planning model for a wafer foundry, which has several existing fabs but needs to construct a new fab. Recent studies aim to solve the problem for the cases of uncertain demand. Yet, most of them do not involve cycle time constraint. Cycle time, a key performance index in particular in semiconductor foundry, should not be ignored. In the first topic, the uncertain demand is modeled as a collection of scenarios. Each scenario, with an occurrence probability, represents the aggregate demand volume under a given product mix ratio. In such a scenario, the mean cycle time of products should be under a predefined target, and the objective of the planning is to maximize the amount of profit. In the second topic, the model formulates and solves a tooling problem in the context of multi-product mix, where the mean cycle time must be under a predefined target. This topic is motivated by the fact that a wafer foundry frequently faces the need to manufacture in various product-mix due to season factors. In the third topic, we describe a tool planning model for constructing a new fab in a company with multiple existing fab sites. Each existing fab originally runs for a particular product mix. To minimize the tool investment for new fab, the company needs to re-allocate the demand to each fab. We present an integrated approach to determine the optimal demand mix and the associated tool plan for the new fab, which can minimize the tool cost of the new fab while each fab (new or existing) is requested to meet a predefined target in its mean cycle time. This dissertation proposes a genetic-algorithm based solution methodology embedded with a queuing analysis to solve the problem. Test examples reveals that the proposed methods greatly outperform the other planning approaches.