爐管製程參數最佳化之研究

Abstract

在半導體爐管多晶矽摻雜製程中,由於該製程複雜且非線性之化學與物理反應，導致產品阻值變異較大,必需時常針對該製程進行參數微調在過去,製程參數微調必需仰賴經驗豐富的工程師,但製程參數的設定無法長期依賴工程師的經驗與直覺。因此，本論文提出一個以類神經網路為基礎的方法，經由倒傳遞神經網路(Back Propagation Neural Network)訓練與測試，用以構建製程參數預測模式。再搭配全數搜尋法,分別以設定溫度與沈積時間,依所需之阻值目標找出最佳製程參數組合。

In semiconductor manufacturing, the poly doped diffusion process is designed to produce a layer of thin film. Due to complex physical and chemical reactions, the resistance of the thin film varied dynamically. Frequent adjustments of process parameters are therefore needed. In practice, the decision of such process parameter adjustments was based on a simple linear interpolation technique, which is not very effective and leads to a high variation on the film resistance. To reduce the variation of film resistance, this research used the technique of back-propagation neural network (BPNN) and developed several predictor models for determining process parameters for the next run. The development of these predictor models is based on a set of sampled data. And of these predictor models, the one that considers the manufacturing information of the last three runs performs with the best accuracy and is called the best-practice model. Based on a large amount of production data, we could justify that the best-practice model is more effective than the traditional linear interpolation technique in reducing the variation of film resistance.