Process Window Qualification (PWQ) is a well-established procedure in semiconductor manufacturing for determining the Process Window (PW), namely the range values of focus and exposure dose that a lithographic process can withstand without deforming the product. The PWQ is performed in practice by printing chips at different values of focus and dose, which are captured by a Scanning Electron Microscope (SEM) at selected locations and then an engineer compares each SEM image against a defect-free reference to assess whether such SEM image presents defective areas. The presence of defective regions in a SEM image indicates that the corresponding focus and dose modulation has to be rejected from the PW. The review represents a tight bottleneck in the procedure, since several images, up to thousands, needs to be inspected. In this thesis we move a step toward the full automation of the PWQ in the semiconductor industry, by addressing the review phase as a reference-based image segmentation problem. This allows to adopt Data-driven Computer Vision techniques to automatically detect and locate defects in SEM images. In particular, we design a Convolutional Neural Network which operates patch-wise and assesses the similarity between each patch and the corresponding reference one to determine whether it contains defects or not. During inference, our CNN processes images in a fully-convolutional manner and provides a heatmap that highlights the defective regions. Our solution dramatically speeds up the review phase of the PWQ, thus enabling the inspection of additional locations, leading to a better characterized PW. The practical contribution of this project involves the design, development and implementation of a NN-based software pipeline that process images and generates corresponding binary images, named Defect Maps, that denotes the predicted locations of defective circuit in the input. Additionally, a graphical user interface has been developed to generate compatible ground truth examples of Defect Maps, according to supervised learning settings.
Process Window Qualification (PWQ) è una procedura ben consolidata nell'ambito dell'industria dei semiconduttori al fine di determinare la Process Window (PW o finestra di processo), la quale si definisce come l'insieme di valori di fuoco e dose di esposizione a cui il processo litografico può sottostare senza causare deformità nel prodotto. La PWQ viene eseguita in pratica stampando chip a diversi valori di dose e fuoco, i quali vengono poi ispezionati da un Scanning Electron Microscope (SEM o microscopio a scansione elettronica) a certe coordinate così che un operatore possa comparare ogni immagine SEM con il suo riferimento non difettoso, in moda da determinare se l'immagine ispezionate presenta aree difettose. La presenza di regioni difettose in un'immagine SEM indica che la corrispondente modulazione di dose-fuoco deve essere rifiutata dalla PW. La revisione causa un forte rallentamento nella procedura, dato che le immagini da ispezionare possono essere diverse migliaia. In questa tesi muoviamo un passo verso la completa automazione della PWQ nell'industria dei semiconduttori, affrontando la fase di revisione come un problema di reference-based image segmentation. Questo permette di adottare tecniche Data-driven di Computer Vision al fine di individuare e localizzare automaticamente difetti in immagini SEM. In particolare, progettiamo una rete neurale convoluzionale (CNN) che opera a livello di patch e calcola la similarità tra ogni patch ed il relativo riferimento al fine di identificare eventuali difetti. Quando applicata, la nostra CNN processa immagini in maniera Fully-Convolutional e fornisce una mappa che evidenzia le regioni difettose. La nostra soluzione velocizza drasticamente la fase di revisione della PWQ, permettendo così l'ispezione di più regioni, migliorando la caratterizzazione della PW. Il contributo effettivo di questo progetto consiste nella ideazione, sviluppo ed implementazione di un software basato su rete neurale che processa immagini e genera corrispondenti immagini binarie, chiamate Defect Map, che denotano le presunte regioni di difettosità nell'input. Inoltre, è stata sviluppata un'interfaccia grafica che permette di generare esempi di Defect Maps, come previsto negli scenari di apprendimento supervisionato.
Toward automatic process window qualification in semiconductor manufacturing through deep learning
MARTINI, MICHELE
2019/2020
Abstract
Process Window Qualification (PWQ) is a well-established procedure in semiconductor manufacturing for determining the Process Window (PW), namely the range values of focus and exposure dose that a lithographic process can withstand without deforming the product. The PWQ is performed in practice by printing chips at different values of focus and dose, which are captured by a Scanning Electron Microscope (SEM) at selected locations and then an engineer compares each SEM image against a defect-free reference to assess whether such SEM image presents defective areas. The presence of defective regions in a SEM image indicates that the corresponding focus and dose modulation has to be rejected from the PW. The review represents a tight bottleneck in the procedure, since several images, up to thousands, needs to be inspected. In this thesis we move a step toward the full automation of the PWQ in the semiconductor industry, by addressing the review phase as a reference-based image segmentation problem. This allows to adopt Data-driven Computer Vision techniques to automatically detect and locate defects in SEM images. In particular, we design a Convolutional Neural Network which operates patch-wise and assesses the similarity between each patch and the corresponding reference one to determine whether it contains defects or not. During inference, our CNN processes images in a fully-convolutional manner and provides a heatmap that highlights the defective regions. Our solution dramatically speeds up the review phase of the PWQ, thus enabling the inspection of additional locations, leading to a better characterized PW. The practical contribution of this project involves the design, development and implementation of a NN-based software pipeline that process images and generates corresponding binary images, named Defect Maps, that denotes the predicted locations of defective circuit in the input. Additionally, a graphical user interface has been developed to generate compatible ground truth examples of Defect Maps, according to supervised learning settings.File | Dimensione | Formato | |
---|---|---|---|
2021_04_Martini.pdf
solo utenti autorizzati dal 08/04/2022
Dimensione
8.09 MB
Formato
Adobe PDF
|
8.09 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/173720