Abstract
Semiconductor manufacturing process is one of the most complicated production environments owing to the challenges of dynamic job arrival, job re-circulation, long production length, and bottleneck drifts. Traditional production planning methodologies were limited to estimate the corresponding throughput and cycle time under various WIP levels with uncertain factors and production constraints. This study aims to develop a data mining framework for cycle time prediction with input factors of production line status such as WIP, capacity, utilization, etc, combining with domain knowledge, to derive empirical rules that the levels of input production factors can be controlled to thus control cycle time and throughput. This approach integrated several data mining techniques in this two-phase research framework including self-organizing maps, decision tree analysis, back propagation neural network, and Gauss-Newton nonlinear regression method. We conducted an empirical study in which real production line data from a semiconductor fabrication factory in Hsinchu Science Park are collected for validation of this framework. The forecast results showed that the proposed framework can derive the productivity performance curves with low forecast error most of times; even sometimes the fab productivity change violently, the forecast models can still obtain an effective result to decrease the forecast error and re-align the forecast models immediately in a few days.