晶圓製造完成後，會針對每片晶圓做完整的晶圓良率及不良狀況測試，確定是否可達成客戶或是公司的期望規格，並且可以提供評估晶圓品質好壞的重要指標，以了解晶圓廠製程品質的穩定性，同時為了提昇測試良率，往往會重測特定不良問題之晶圓。 一般來說，不良 晶圓的分類方法，是透過 Bin 數字的分類 來檢測晶圓上各種不同不良的指標，在分析的過程中，首先測試工程師必須先分析良率是否符合目標值，如果晶圓良率低於目標值，可考慮針對特定不良晶圓執行重測動作。由於不良的原因可能來自於晶圓廠製程異常、測試設備不穩定，或是因測試作業不當而造成；然而大部分工程師所關心的，除了必須分析造成良率異常的原因外，更想了解在重測後晶圓良率是否可提高，期望在重測後的良率比未重測前能明顯提高，因此為了提升晶圓良率，往往只注意良率的變化，而未考量實際所帶來的重測效益。 本研究收集案例公司於 95 年某產品的晶圓測試資料、測試成本與產品利潤之相關資料以供研究分析，論文首先探討案例中影響不良的原因為何，並透過基因演算法尋找最適之重測門檻值，符合最佳的測試流程，也希望為案例公司帶來最佳效益，為此論文之主要貢獻。 實驗的分析顯示，所提出的基因演算法方法得到的重測門檻值最後得到的利潤為13182474，比實際個案公司的重測門檻得到的利潤13144288還多出38186的改進，若是只看需要重測的晶圓，則利潤為2711771.2，比原個案公司的2365971.8改進了114.62%，因此可以說由此研究提出的基因演算法求解晶圓重測門檻值有不錯的結果。 In order to verify whether wafer can be able to achieve expected specification, generally every die will be probed completely after being fabricated in wafer fab. By such operation, we can use the essential index to improve wafer foundry’s quality and also the yield rate of wafers. In the mean time, IC design house hopes to get better yield rate during wafer probing process, normally they will try to retest low-yield-wafer, no matter retest entire gross dies or particular defective dies. In general, defective dies are classified by using different bin numbers; the bin numbers represent particular testing result or its performance. During failure analysis after finishing wafer testing, testing engineer can decide to re-test abnormal wafer directly if it’s out of yield limit set previously. As normally engineers just ‘hope’ to get higher yield recovered from second testing, they seldom know how to predict yield variation and regard the related profit before making decision of re-testing Hence, this thesis attempts to propose a workable solution for wafer testing process by using Genetic Algorithm to establish thresholds of wafer retesting. Through a real example from CMOS Image Sensor probing process, it was presented to demonstrate the methodology. The result of experiment show that the profit of this study is 13182474, and it is greater than the profit of case company, increases 38186. If we only discuss the retesting wafer, the profit is 2711771.2, compares with case company, and its improvement is 114.62%. So that could say Genetic Algorithm has a not bad solution with threshold of wafer retesting in this thesis.
