| 摘要: | 半絕緣碳化矽(semi-SiC)晶圓是第三代半導體材料之一，由於其具有極高的硬度與脆性，傳統加工方法在該材料製作複雜的三維幾何結構時面臨嚴峻挑戰。放電加工(EDM)被視為加工此類材料的一種可行替代方案。本論文探討了semi-SiC晶圓之放電加工微鑽孔與線切割放電加工(WEDM)邊緣修整技術，並應用輔助導電層技術以實現對非導電材料的有效加工。
本研究分為兩個主要部分。第一部分著重於建立深度神經網路(DNN)模型，以預測EDM微鑽孔之加工品質，並以入口孔徑與出口孔徑作為品質指標。研究中使用隨機森林回歸(RFR)分析脈衝開啟時間、脈衝關閉時間、峰值電流與加工時間主要加工參數的重要性。並透過反應曲面法(RSM)評估參數間之交互作用。此外，採用粒子群演算法(PSO)以最佳化DNN結構，在訓練與測試階段皆展現出高度的預測準確性。模型的準確性表現出較低的平均絕對百分比誤差(MAPE)、均方誤差(MSE)、均方根誤差 (RMSE)，以及較高的決定係數(R2)上。本研究所建構結合PSO的DNN模型，成功找出最佳加工參數，並經由實驗結果驗證其有效性，加工品質得到顯著提升。同樣地，本研究亦探討了WEDM修整過程中修整半徑(R)與表面粗糙度(SR)之最佳化問題，所考慮的加工參數包括開路電壓、脈衝開啟時間、脈衝關閉時間與伺服電壓。根據81組全因子實驗設計結果，WEDM修整可實現低於0.5 μm的表面粗糙度，符合商業加工標準。此外，亦建立一個含六層隱藏層之DNN模型，並整合PSO，可用以精準預測與最佳化R與SR品質，成功獲得較佳之加工參數。
第二部分對加工參數對semi-SiC晶圓之表面與次表面特性的影響進行探討，特別著重於EDM微鑽孔之脈衝開啟時間與峰值電流，以及WEDM修整之開路電壓與脈衝開啟時間。本研究採用掃描式電子顯微鏡(SEM)、能量色散X-射線光譜(EDS) 、X光光電子能譜儀(XPS)與穿透式電子顯微鏡(TEM) 分析技術，觀察加工表面因熱能作用所產生之坑洞、重凝固物、裂縫與微孔等特徵。EDM之材料移除機制包含熔融、汽化、脫落與氧化。EDS分析結果顯示，隨著放電能量提高，加工表面之碳與氧含量顯著增加，此現象可歸因於EDM加工過程中SiC的分解與氧化反應。XPS分析進一步確認加工表面存在石墨、二氧化矽(SiO2)、銅粒子、氧化亞銅(Cu2O)及氧化銅(CuO)等物質。TEM分析進一步識別出重鑄層、熱影響區(HAZ)以及未受影響之SiC本體，其層厚會隨放電能量變化而有所不同。相較於EDM微鑽孔，WEDM修整因其放電能量分布較為分散，且總放電能量較低，得以生成較薄的重鑄層與熱影響區。綜合上述結果顯示，相對於其他加工方式， EDM微鑽孔與WEDM可有效降低表面與次表面損傷，有助於維持semi-SiC晶圓之結構完整性，具潛力應用於高精密製程。
;Semi-insulating silicon carbide (semi-SiC) wafers, recognized as one of the third-generation of semiconductor materials, exhibit exceptional hardness and brittleness, making the machining of complex three-dimensional geometries challenging by using conventional methods. Electric discharge machining (EDM) offers a promising alternative for processing such materials. This dissertation investigated micro-hole EDM drilling and wire EDM (WEDM) trimming of semi-SiC wafers, employing an assisting electrode technique to enable the machining of non-conductive materials.
This study is divided into two main parts. The first part focused on developing a deep neural network (DNN) model to predict machined quality in micro-hole EDM drilling, with inlet and outlet hole diameters selected as quality indices. Major machining parameters, including pulse-on time, pulse-off time, peak current, and working time, were analyzed using Random Forest Regression to assess parameter importance and Response Surface Method (RSM) to evaluate parameter interactions. Particle Swarm Optimization (PSO) was employed to optimize the DNN architecture, achieving high predictive accuracy as indicated by low mean absolute percentage error (MAPE), mean squared error (MSE), root mean squared error (RMSE), and a high coefficient of determination (R²) during training and testing. The developed DNN model, coupled with PSO, successfully identified optimal machining parameters, which was validated through experiments that demonstrated improved machined quality. Similarly, WEDM trimming was investigated to optimize trimmed radius (R) and surface roughness (SR), considering machining parameters including open voltage, pulse-on time, pulse-off time, and servo voltage. Based on a full factorial experiment of 81 trials, the WEDM trimming achieved sub-0.5 μm SR, meeting commercial standards. A six-hidden-layer DNN model, also integrated with PSO, was developed and could effectively predict and optimize R and SR, outperforming initial experimental results.
The second part examined the effects of machining parameters on the surface and subsurface characteristics of semi-SiC wafers, considering pulse-on time and peak current in micro-EDM drilling, and open voltage and pulse-on time in WEDM trimming. Scanning electron microscopy (SEM), energy dispersive X-ray spectroscopy (EDS), X-ray photoelectron spectroscopy (XPS), and transmission electron microscopy (TEM) were employed for microstructural characterization. Craters, resolidified material, cracks, and micro-pores were found on machined surfaces, as driven by thermal energy. Material removal mechanisms involved melting, vaporization, spalling, and oxidation. EDS analyses showed increased carbon and oxygen content with rising discharge energy, attributed to SiC decomposition and oxidation during EDM. XPS analysis indicated the presence of graphite, SiO2, copper particles, Cu2O, and CuO on machined surfaces. TEM analysis identified a recast layer, heat-affected zone (HAZ), and unaffected bulk SiC, with layer thickness varying with discharge energy. Compared to micro-EDM drilling, WEDM trimming produced thinner recast layers and HAZ due to lower, more dispersed discharge energy. These findings highlighted the advantages of micro-EDM and WEDM in minimizing surface and subsurface damage, thereby preserving the structural integrity of semi-SiC wafers for high-precision applications. |
