Page 59 - Vol.47

TSMC / FACILITY PUBLISHED
VOL.47
廠務季刊

URL.http://nfjournal/

應用AI-AMC機械學習預測與污染來源分析

AI application-AMC prediction by machine learning and pollution source analysis

文│莊鎧瑋│Arizona Facility│

摘要 / Abstract

因應市場的需求，半導體製程技術演進快速，電晶體30年內已縮小了1000倍，到了五奈米的技術，而隨著規格及產線環境
要求的提升，氣態分子微汙染(airborne molecular contamination, AMC)的控制對於製程良率改善更是有舉足輕重的影響。AMC
汙染源普遍來自製程清潔使用後化學品，煙囪排放後由外氣空調箱進入FAB，造成潔淨室環境汙染濃度上升而頻繁更換化學濾
網，進而衍生龐大的濾網相關費用。AMC的防治及監測更是需投入大量的物資、人力，因此若能利用現有的資料庫，對多面項
數據進行結合、比較甚至預測，將有效提升AMC的防治。
隨著大數據及機器學習的蓬勃發展，本文利用機器學習模型和AMC大數據的結合進行汙染的來源分析並改善，再進一步預
測未來24小時內MAU出口的AMC濃度變化，猶如天氣預報，作為及早應對措施之參考，減少警報次數甚至降低運轉成本。

關鍵詞 / 氣態分子微汙染、潔淨室、機器學習、大數據

Over the past 30 years, the semiconductor technology has been developing swiftly and exponentially in response to the
skyrocketing demand of the market, the size of transistor has shrunk by 1,000 times and delved into five-nanometer scale. With
progressively stringent production-line environmental and specification requirements, gaseous molecular pollution(airborne
molecular contamination, AMC) control plays a paramount role in the yield enhancement. The source of AMC typically originates
from chemicals(solvents) used for cleaning wafers in the processes, after being discharged from the stack, through the air
conditioning box into the FAB, resulting in increasing contaminant concentration in the clean room and frequent replacement of
AMC chemical filters, which in turn leads to massive cost associated with the filters. The precautions and monitoring of AMC require
a tremendous amount of materials and manpower. Therefore, effective control of AMC can be achieved with early preventive
measures if the existing database can be combined, compared, and even utilized to predict the trend of AMC down the road.
With the mature development of big data analysis and visible advancement of machine learning, this study used the
combination of machine learning model and big data to conduct analysis and prediction, furthermore, foresee the change of
the AMC concentration at the MAU discharge in the following 24 hours, just like weather forecast, as a reference for counter-
measures to reduce the alarm rate and operating costs.

Keywords / Airborne Molecular Contamination, Cleanroom, Machine Learning, Big Data

1. 前言

隨著半導體製程的演進，電晶體規格從1987年的3微米到 霧化(MA/MC) [14][15] 、電性的改變(MD)等產品的缺陷。因此，工
2021年的3奈米，無塵室生產環境對於AMC潔淨度的要求亦日 廠針對AMC汙染物的防治，每年投入大量成本於監測系統的
漸嚴苛。許多文獻指出AMC對於半導體製程的重要影響，如 設置、保養，人力巡檢及採樣，濾網的購買、更換及處理。而
造成導線的腐蝕(MA) 、T-topping(MB, NH 3 ) 、微影鏡頭的 透過儀器監測，人工採樣提供大量的AMC數據基礎，若能藉
[13]
[12]
FACILITY JOURNAL 09 2022 57