Neural network based process control of liquid crystal displays (LCDs) manufacturing line

Abstract

TFT-LCDs(Thin Film Transistor-Liquid Crystal Displays) are promising candidates for the next generation display terminals since they are flat and light-weight. The market for these large sizeLCDs with TFT is expected to expand dramatically near future. The yield improvement is the most important theme for TFT-LCDs production because it leads to cost reduction. However, large-area TFT-LCDs have a large critical area and are defect-sensitive. This make yield production difficult. To solve this problem, yield estimation is very effective. For TFT design, yield cannot be forecasted before fabrication by the estimation. Futhermore, for yield management, stricter process control can be realized by establishing control limits by statical estimation. Many processing steps are required to manufacturing a TFT, providing ample opportunity for the introduction of various processing defects that the reason of low yield, causulity of defect and correlations among process parameters. The neural network methodology as a diagnosis for TFT-LCDs applied to obtain more accurate relation between the unit processing control level and TFT array yield. The sensitivity analysis can be also applied to a trained neural network to extract useful information which maybe difficult to obtain through traditional analytical methods, such as what are the most important variables with respect to the TFT manufacturing yield. We suggest the heuristic search methodology, as a Neuro process simulation, for optimal control of main deposition and sub-process in determining the cause of the deviation of unit process control range, and provide a controlling basis for us in order to operate the plant more efficiently.