- 學位論文
應用人工智慧於高階IC誤宰測試研究
Utilizing Artificial Intelligence for Advanced IC Overkill Detection Research
摘要
半導體製造是由多個製程階段組成的加工過程，每個階段都需要先進的資料解析工具來提高營運效率並降低生產成本。IC晶片製造過程主要包括晶圓設計與製造、晶圓測試(Chip Probe, CP)、封裝(Packaging)和成品測試(Final Test, FT)。部分產品會在製程中增加模組測試(System Level Test, SLT)以確保產品功能品質。然而，過去文獻多數都是以晶圓製造或晶圓測試的製程良率作為研究對象，很少研究涉及以成品測試製造業的觀點，使用分析工具協助處理測試誤宰(Overkill)與無效重工(Rework)的檢測，降低誤宰率且避免測試座的無效清潔。本研究以人工智慧的羅吉斯迴歸(Logistic Regression)、支持向量機(Support Vector Machine, SVM)、隨機森林(Random Forest)技術方法，解析成品測試所產生的測試資料，檢測測試誤宰，從而避免生產之無效測試產出和測試不良品重工狀況的發生。 實證分析以新世代專業測試廠X公司實際量產的TFBGA 591 22.8mm x 22.5mm高階系統晶片(System on Chip, SoC)的測試資料作為人工智慧－機器學習的資料來源。研究擷取了相同機台號(93K#20)於2023年7月28日至2023年8月3日的連續十個批次，共計53,226顆成品的測試資料，並將首次測試與重測測試資料全部納入。隨機森林的結果顯示準確率(Accuracy score)達到99.56%，積極指標(Recall score)同樣為99.56%，保守指標(Precision score)為86.93%，調和平均值(F1 score)為91.88%。這些結果顯示使用人工智慧檢測IC成品測試誤宰準確性方面具有顯著優勢。另外，隨機森林的訓練時間為118.52秒，測試時間為1.17秒，本研究得出隨機森林模型不僅具有高效的檢出能力，還具備快速的執行效率。 本研究為成品測試業中的成品測試製程提供了一套完善的生產監控與管理策略，有助於進一步提升行業競爭力，同時提高了成品產品測試的精度與效度。
並列摘要
Semiconductor manufacturing is a complex process composed of multiple stages, each requiring advanced data analysis tools to improve operational efficiency and reduce production costs. The IC chip manufacturing process mainly includes wafer design and manufacturing, wafer testing (Chip Probe, CP), packaging, and final testing (Final Test, FT). Some products add system-level testing (System Level Test, SLT) during the process to ensure product functionality and quality. However, most previous literature focuses on yield rates in wafer manufacturing or wafer testing processes, with little research addressing the perspective of final test manufacturers. This study employs analytical tools to assist in detecting overkill and ineffective rework, reducing the overkill rate, and avoiding unnecessary cleaning of test fixtures. We use artificial intelligence techniques, such as Logistic Regression, Support Vector Machine (SVM), and Random Forest, to analyze the test data generated during final testing, predicting test defect trends and issuing early warnings to avoid ineffective test outputs and rework of defective products. The empirical analysis uses test data from TFBGA 591 22.8mm x 22.5mm high-end System on Chip (SoC) products, produced by a new-generation professional testing company, X Company, as the data source for AI machine learning. The study extracted test data from ten consecutive batches, totaling 53,226 finished products, tested on the same machine (93K#20) from July 28, 2023, to August 3, 2023. All first tests and retests were included. The results show an accuracy score of 99.56%, a recall score of 99.56%, a precision score of 86.93%, and an F1 score of 91.88%. These results indicate that Random Forest significantly improves the accuracy of IC final product testing. Additionally, the training time for the Random Forest model was 118.52 seconds, and the testing time was 1.17 seconds, demonstrating that the model not only has high detection capabilities but also rapid execution efficiency. This research provides a comprehensive production monitoring and management strategy for final testing processes in the finished product testing industry, helping to further enhance industry competitiveness while improving the accuracy and validity of final product tests.