本研究首先探討彩色濾光片微影製程的相關文獻並與工程師討論，決定產品品質特性與控制參數。再藉由田口實驗規劃及共變異數分析，將多品質目標的問題，轉化整合成單一品質目標的問題。資料分析包括因子反應分析及變異數分析等，據此瞭解實驗控制因子對S/N比與品質特性之間的關係。其次利用S/N比因子反應圖，可找出對品質特性影響較顯著之重要控制因子。但是由於田口實驗設計規劃所得的最佳解仍屬於離散型的最佳解，而且所得之最佳解未必能夠同時符合製程穩定與品質目標，因此本研究使用改良式混合粒子群與基因演算法(hybrid PSO-GA)結合多層感知器，找出倒傳遞類神經網路之初始權重，針對類神經網路權重最佳化，並將最佳化的權重值作為倒傳遞類神經網路的初始值進行訓練與測試，藉此改善倒傳遞類神經網路訓練的速度與精準度，建構出S/N比預測器與品質特性預測器，並以S/N比預測器與品質特性預測器結合改良混合粒子群最佳化與基因演算法，利用變異數分析找出製程顯著之控制因子，以田口實驗設計分析而得的參數組合作為初始值，進行數值分析，找出能夠同時符合製程穩定以及品質目標的最佳製程參數組合，並針對製程參數與品質特性之關係進行探討。本研究之智慧最佳化系統不僅提升彩色濾光片微影製程品質同時能有效提升製程穩定。

This study integrates the Taguchi method, analysis of variables (ANOVA), Multilayer Perceptron (MLP), back-propagation neural networks (BPNN), and modified hybrid PSO-GA to develop a multi-objective intelligent parameter optimization system of the photolithography process for color filter. The quality characteristics of product and control variables can be well-ascertained, then transforming the problem from multi-objective quality characteristics combing into a single quality characteristic using the experimental planning of Taguchi method and covariance analysis. Moreover, identifying the relationships between control factors of experiments and signal-to-noise (S/N) ratios and quality characteristics through the data analyses comprising main effects and interactions of factors and analysis of variables (ANOVA). As a result, the crucial control factors can be found by main effects and interactions graphs of factors in terms of signal-to-noise (S/N) ratios, and the better Taguchi parameter settings can also be obtained through the sorting of control factors in the multi-quality characteristics situation. However, the optimal parameter settings (solution) through the Taguchi experimental planning is still belong to a discrete optimal solution, and unnecessarily meet the process stability and quality goals. Therefore, this study identifies the initial weight of back-propagation neural networks (BPNN) using modified hybrid PSO-GA with multilayer perceptron (MLP) to improve the BPNN training efficiency and precision by regarding the optimal weight as BPNN initial value, and then implementing to train and test the BPNN; finally, the study constructs the signal-to-noise (S/N) ratios and quality characteristics predictors combining with modified hybrid PSO-GA, and analysis of variable (ANOVA) to locate the significant control factors, serves the parameter settings of Taguchi experimental design as the initial value, proceeds to analyze the data, and further pinpoint the optimal parameter settings which can achieve the quality goal and reach the process stability at the same time. Accordingly, the proposed intelligent optimization system can not only enhance the product quality but also make the process stability in effect.