In semiconductor manufacturing, the process of short-term production planning requires setting clear and yet challenging and doable goals to each operation and toolset in the process flow per each product type. We demonstrate the complexity of this problem using an experimental study performed with proficient workforce, and then show how the problem can be decomposed, aggregated, and solved using sequential recurrent linear programming assignment problems. We also refer to the improvements that the proposed algorithm has achieved in practice when applied to multiple semiconductor production facilities, and discuss its efficiency and uniqueness as a fast heuristic relative to other proposed methods.
Today, in the semiconductor industry, more and more fabs run multiple processes with a diverse range of products (Bang, An, Kim, & Lim, 2005). Since a majority of the resources (equipment and head-count) are shared amongst the various products, the need for short-term production planning to efficiently meet on-time delivery and output goals becomes a complex continuous process of resource allocation planning for each toolset (or machine group) throughout all processes, operations, and products/lots (Zorea, Perez, Pridor, & Bregman, 2003).
In the late 1990s, drum-buffer-rope (DBR) based policies were attempted at the problem, which proved to work very well for regular production lines and systems. For example, Sivasubramanian, Selladurai, and Rajamramasamy (2000) report on a successful case study that has been taken up in a small-scale industry, and analysis has been carried out on the positive effect of the DBR approach on the performance of the system. However, even without the complexity emanating from the product mix environment, the highly re-entrant process in semiconductor manufacturing systems dictates that the same machines perform competing (different) operations on the lots as they are processed through the fab and consequently, implementing drum-buffer-rope is difficult and complex in re-entrant flows, because several bottleneck operations of each lot will appear on the drum in various locations (Wu & Yeh, 2006).
A number of specific dispatching rules have been widely applied to determine the best strategy of machine allocations to lots in process at the time of execution, such as 3–2–1, also known as Back-To-Front (Sohn & Kempf, 1996) and Critical Ratio (Subramaniam, Ming, & Mohan, 2005). A combination of these dispatching rules has also been evaluated (Dabbas & Fowler, 2003). However, attempting to solve an underlying short-term planning problem implicitly with dispatching rules, at factories with a range of several technologies, multiple products per technology, and hundreds of operations in the process flow – is almost set to failure, since these rules are by definition limited in the consideration of pertinent information for making the best decision.
In this paper, we illustrate the complexity of the short-term production planning decision problem, using a re-entrant line consisting of only two products and eight operations over a planning horizon of 1 week (fourteen shifts), where performance is measured on output and on-time delivery. We report the results of an experimental study that was performed with this line, with many proficient participants from across the semiconductor manufacturing industry that deal with this problem on a daily basis. The experimental study evaluates the goodness of solutions based on common practices such as 3–2–1, critical ratio, fast-box and others in obtaining near-optimal solution to the problem.
We then compare these results with those obtained by a unique algorithm proposed for this problem. The algorithm is based on decomposition of the short-term production planning problem into a sequence of assignment sub-problems (of available machines to available lots), each solved to optimality by itself and, when aggregated together, form a comprehensive heuristic solution.
Similar to the approach by Qiu (2005), we present a practical solution to the problem while also addressing issues such as real-time performance, scalability, and re-configurability. Unlike in Qiu, which is based on distributed WIP control, the proposed algorithm is a fast and efficient heuristic that has easily been applied in real-time for large-scale semiconductor manufacturing facilities. The proposed algorithm has been demonstrated via multiple implementations in high volume production facilities, and we also reference the improvements that this algorithm has achieved in practice once it was implemented in these facilities.
In this paper, we illustrate the complexity of the short-term production planning decision problem, using a small-scale re-entrant line.
Various complexity factors make this short-term production planning to efficiently meet on-time delivery and output goals become a complex continuous process of resource allocation planning for each machine group throughout all product–operations, and WIP. Thus, as demonstrated in our experimental study, humans cannot comprehend the pertinent data, arising conflicts, and multiple objectives and constraints altogether and tend to make mistakes on the way. On the other hand, a fast heuristic algorithm that is based on a solution for sequential time-discrete recurrent assignment problems, such as the one outlined in this paper, proves to be very efficient in this case.
The practicality of this solution approach to the problem is not only by means of efficiency but also in the ability to apply it in real-time and for large-scale problems (as described in a separate paper). The scalability and re-configurability properties of the proposed approach without compromising on a centralistic WIP control, make it highly viable for the high volume high mix semiconductor manufacturing environment. The proposed algorithm has been tested successfully in high volume production facilities, yielding improvements of 5–6% in shiftily moves, and even a larger improvement of 12–13% in levels of in-process (indicative of more efficient tool utilizations). In addition, a reduction in the output variability was observed from a typical 12–13% standard deviation in shiftily output to 5.5%, on the average, after the implementation. WIP-Turns (measure of line velocity) increased by 20% on the average and up to 100% for low volume products. Delivery performance was also significantly improved to from around 90% to >99% over 30–50 product-types.
