對於使用機器學習技巧預測臨界電壓之測試時間縮減流程

Abstract

傳統的晶圓接受度測試需要很長的測試時間才能量測出待測元件的臨界電 壓變異程度，在先前的研究已經提出一個新的測試架構，搭配模型擬合技術，可 以快速且有效的計算出大量待測元件的平均與變異數。本論文更進一步的將此測試結構加大以增加預測得精準度。接著，本論文更提出一個修改過後的模型擬合流程，目的在於再更進一步的降低計算臨界電壓變異程度所需要量測電流的次數，此流程成功的將量測次數從原本的21 次降至3 次。在論文的後段，更嘗試使用這套流程去預測一個新的目標，線性區的臨界電壓，預測平均數與變異數的結果亦能達到超過99% 的決定系數。而相較於傳統使用二元搜尋法的接受度測試，整體的測試時間可以達到228.3 倍的加速。 i

Conventional array based test structure require long measuring time to measure the variation of device Vt. Previous work we have proposed a new array-based test structure which can obtain mean and variance of Vt in shorter measuring time with a model-fitting framework. In this thesis, we further extend the DUT number to increase the accuracy for predicting. Then we revised our previous low to reduce the total Vg we used from 21 to 3, which can further reduce the overall test time required. At last, we try to predict a new value, variation of linear Vt with our revised model-fitting flow and achieve more than 99% R2 for predicting its mean and variance. Compare to conventional method, we have 228.3x speedup in overall test time.