| 摘要: | 近年來第三代半導體飛速發展，以其耐高溫、高功率的特性在電動車、電源供應器等領域大放異彩，能夠在極端的環境下仍持續運作，且能簡化電子產品的電路設計與散熱系統等。但第三代半導體的代表材料之一的碳化矽價格較高昂，由於碳化矽材料需要嚴苛環境進行長晶，且碳化矽的長晶速度緩慢，需要經過大量加工程序才能投入使用，造成碳化矽晶圓生產不易。為了降低碳化矽材料成本，必須優化晶圓製造技術。線切割放電加工是一種成熟且高精度的加工技術，擁有多種可控參數可以針對目標導體材料進行調整，透過更改部分參數可以改良工件的加工品質，且在加工過程也可以即時監測間隙電壓等動態參數。為了最佳化線切割的加工參數並分析參數配置對碳化矽加工品質之影響，本研究透過深度學習的方式改良加工過程，模型使用長短期記憶神經網路(Long Short Term Memory, LSTM)進行預測，LSTM 模型擅長學習時間序列等資料並分析時間序列的因果關係特徵，蒐集線切割加工過程中的參數變化時序作為輸入，模型能夠有效分析線切割的動態參數變化對加工品質的影響。最後面對碳化矽樣本數量較少、蒐集不易的問題，本研究使用遷移學習方法搭配模型共同訓練。遷移學習方法能夠降低模型訓練所需的碳化矽學習樣本數，透過大量相似的矽晶圓樣本配合碳化矽樣本進
行訓練，可以改善模型的泛化程度、解決過擬合等問題並充實模型的訓練集。本研究的目的為建立線切割品質預測模型，利用儀器蒐集加工過程的時序資料，分析各加工參數對碳化矽晶圓之表面粗糙度與切口寬度的影響，以 LSTM 模型配合三種遷移學習方法建立模型，輸出樣本的表面粗糙度與切口寬度。本實驗的輸入為兩種動態參數時序資料與三種固定參數組合而成，固定參數共擁有 27 種參數配置，以矽晶圓作為源域資料集共有 270 個樣本，碳化矽作為目標資料集共有 270 個樣本，訓練集與驗證集比例為 6：4。驗證集之預測結果，表面粗糙度與切口寬度之 R-square 值分別為 95 與 82左右由此可見本模型能夠有效預測碳化矽之加工品質，減少加工過程與深度學習上的材料成本。;In recent years, the third generation of semiconductors has developed rapidly. With its high temperature resistance and high power characteristics, it has good effect in the fields of electric vehicles and power supplies and it can operate in extreme environments ,simplify the circuit design ,design of electronic products and Cooling system, etc. However, SiC, one of the representative materials of the third generation semiconductor, is relatively expensive. Since SiC material requires a harsh environment for crystal growth, and the crystal growth rate of SiC is slow, it requires several number of processing procedures before it can be put into use, resulting in SiC Wafer production is difficult. In order to reduce silicon carbide material costs, wafer manufacturing technology must be optimized.Wire Electrical Discharge Machining is a mature and high-precision processing technology. It has a variety of controllable parameters that can be adjusted according to the target conductor material. By changing some parameters, the processing quality of the workpiece can be improved, and the gap can also be monitored in real time during the processing process. dynamic parameters such as voltage. In order to optimize the processing parameters of WEDM and analyze the impact of parameter configuration on the processing quality of SiC, this study improves the processing process through deep learning. The model uses The Long Short Term Memory (LSTM) neural network for prediction. The LSTM model is good at learning time series and analyzing the causal relationship characteristics of the time series. It collects the parameter change time series during the wire cutting process as input. The model can effectively analyze the impact of dynamic parameter changes of the wire cutting on the processing quality. Finally, facing the problem that SiC samples has few samples and difficult to collect, this study uses model training with the transfer learning method. The transfer learning method can reduce the number of SiC learning samples required for model training. Training with SiC samples and a large number of similar silicon wafer samples can improve the generalization of the model, solve
problems such as overfitting, and enrich the model training set.
The purpose of this study is to establish a WEDM quality prediction model, use instruments to collect time-series data of the cutting process, analyze the impact of each processing parameter on the surface roughness and kerf width of silicon carbide wafers and establish LSTM model with two transfer learning methods. The model outputs the surface roughness and kerf width of the sample. The input of this experiment is a combination of two dynamic parameter time-series data and three static parameters. The static parameters have a total of 27 parameter configurations. There are 270 samples of silicon wafer as the source domain data set and 270 samples of SiC as the target data set and the ratio of training set to validation set is 6:4. According to the prediction results of the validation set, the R-square values of surface roughness and kerf width are about 95 and 82 respectively. It can be seen that this model can effectively predict the processing quality of silicon carbide and reduce processing Material costs on process and deep learning. |
