The main products of semiconductor manufacturers are Integrated Circuits (ICs). Fabrication of ICs consists of hundreds of specialized steps which result in a sequence of patterned layers of different materials, gradually formed one on top of the other, on an initial substrate of silicon, called the semiconductor wafer. The overall process is complex, costly and takes up to several weeks. Along all the production line automated inspection tools test a sample of wafers to look for physical defects and to assess production quality. Defects consist of voids, craters, protrusions, bridges, particles, and anything that may cause the chip to fail. The result of this inspection phase is a Wafer Defect Map (WDM), a collection of the defects coordinates (x; y). In normal production conditions, defects are distributed on the wafer surface in a random way, without pointing out geometries or agglomerations. However, when some problem occurs, whether machine or human related, defects display a peculiar spatial arrangement, forming what it is called a pattern or signature. When defective wafers are found, the fabrication may be interrupted and adjustments in the process conditions may be introduced for correction purposes to limit the losses. Identifying some specific defect patterns is a crucial step in analyzing and retrieving the corresponding root causes, and thus being able to take action as soon as possible. Traditionally, the WDMs are visually inspected by human experts to find those causes: operators examine hand-crafted features (for example defect locations, sizes, colors, and shapes) with the help of high-resolution microscopes and can identify the malfunctioning piece of equipment. However, this approach has critical drawbacks: it is highly time-consuming and misidentification or inconsistent classification can occur due to fatigue and human subjectivity. The aim of this thesis project is to build an automated defect pattern classification system that associates to each Wafer Defect Map a class among 12 classes identified in collaboration with STMicrolectronics engineers. The 12 classes are: Normal, Donut, Incomplete, Basket Ball, Cluster Big, Cluster Small, Fingerprints, Geometric Scratch, Grid, Half-Moon, Ring, Zigzag. Normal class corresponds to normal production conditions where no pattern is present; all the other classes correspond to a real pattern and together form the Anomalous class. Our solution takes as input a WDM, determines whether a defect pattern is present and classifies it according to the set of predefined classes. In particular, we cast the automatic classification of WDMs to a image classification problem and we adopt a deep Convolutional Neural Network (CNN) to recognize the defect patterns. Practically, each wafer map is rendered to an image that is passed through the CNN classifier and gets a labeling. To treat the defect pattern classification problem as an image classification, we developed a pipeline to derive for each wafer map the corresponding image. Following this idea, we decided to exploit the well-known effectiveness and promising results of Convolutional Neural Network in image classification.
La fabbricazione di componenti elettronici (IC) consiste in centinaia di passaggi specializzati che danno luogo a una sequenza di strati di materiali diversi, gradualmente formati uno sopra l'altro, partendo da un substrato iniziale di silicio, chiamato wafer. Il procedimento è complesso, costoso e dura fino a diverse settimane. Lungo tutta la linea di produzione strumenti di ispezione automatizzata testano un campione di wafer per cercare difetti fisici e valutare la qualità della produzione. I difetti possono essere vuoti, crateri, sporgenze, ponti, particelle e tutto ciò che può causare il malfunzionamento del chip. Il risultato di questa fase di ispezione è una Wafer Defect Map (WDM), una raccolta delle coordinate (x,y) dei difetti. In condizioni normali di produzione, i difetti risultano distribuiti sulla superficie del wafer in modo casuale, senza evidenziare geometrie o agglomerazioni. Tuttavia, quando si verifica qualche problema, sia che sia dovuto a una macchina o a una persona, i difetti mostrano una particolare disposizione nello spazio, formando ciò che è chiamato un pattern o traccia. Quando vengono trovati wafer difettosi, la fabbricazione può essere interrotta e possono essere introdotti correzioni nelle condizioni di processo a fini di risolvere il problema e limitare le perdite. L'identificazione di specifici pattern difettosi è un passaggio cruciale nell'analisi e nel recupero della causa scatenante, che permette quindi si essere in grado di agire il più presto possibile. Tradizionalmente, le WDM sono ispezionate visivamente da esperti per trovare le cause del difetto: gli operatori esaminano le caratteristiche hand-crafted (ad esempio la posizione dei difetti, le dimensioni, i colori e le forme) con l'aiuto di microscopi ad alta risoluzione e riescono a identificare il pezzo di equipaggiamento malfunzionante. Tuttavia, questo approccio ha diversi inconvenienti: è molto dispendioso in termini di tempo e la scorretta o incoerente classificazione può verificarsi causata da fatica e da soggettività umana. Lo scopo di questo progetto di tesi è quello di costruire un sistema automatico di classificazione di pattern che associa ad ogni Wafer Defect Map una classe tra le 12 classi che abbiamo identificato in collaborazione con gli ingegneri di STMicrolectronics. Le 12 classi sono: Normale, Ciambella, Incompleto, Basket Ball, Cluster Grande, Cluster Piccolo, Impronte digitali, Graffio geometrico, Griglia, Mezzaluna, Anello, Zigzag. La classe normale corrisponde alle normali condizioni di produzione in cui nessun pattern è presente; tutte le altre classi corrispondono a un pattern reale e insieme formano la classe anomala. La nostra soluzione prende come input un WDM, determina se un pattern è presente e lo classifica in base alle classi predefinite. In particolare, consideriamo la classificazione automatica delle WDM come un problema di classificazione di immagini e adottiamo una rete neurale convolutiva (CNN) deep per riconoscere i pattern di difetti. Praticamente, ogni mappa viene trasformata in un'immagine che viene poi passata attraverso il classificatore CNN e viene etichettata. Per trattare il problema di classificazione di pattern come classificazione di immagini, abbiamo sviluppato una pipeline per derivare per ogni WDM la corrispondente immagine. Seguendo questa idea, abbiamo deciso di sfruttare la ben nota efficacia e risultati promettenti della CNN nel compito della classificazione di immagini.
Convolutional neural networks to reveal patterns in wafer defect maps
MOIOLI, LIDIA
2016/2017
Abstract
The main products of semiconductor manufacturers are Integrated Circuits (ICs). Fabrication of ICs consists of hundreds of specialized steps which result in a sequence of patterned layers of different materials, gradually formed one on top of the other, on an initial substrate of silicon, called the semiconductor wafer. The overall process is complex, costly and takes up to several weeks. Along all the production line automated inspection tools test a sample of wafers to look for physical defects and to assess production quality. Defects consist of voids, craters, protrusions, bridges, particles, and anything that may cause the chip to fail. The result of this inspection phase is a Wafer Defect Map (WDM), a collection of the defects coordinates (x; y). In normal production conditions, defects are distributed on the wafer surface in a random way, without pointing out geometries or agglomerations. However, when some problem occurs, whether machine or human related, defects display a peculiar spatial arrangement, forming what it is called a pattern or signature. When defective wafers are found, the fabrication may be interrupted and adjustments in the process conditions may be introduced for correction purposes to limit the losses. Identifying some specific defect patterns is a crucial step in analyzing and retrieving the corresponding root causes, and thus being able to take action as soon as possible. Traditionally, the WDMs are visually inspected by human experts to find those causes: operators examine hand-crafted features (for example defect locations, sizes, colors, and shapes) with the help of high-resolution microscopes and can identify the malfunctioning piece of equipment. However, this approach has critical drawbacks: it is highly time-consuming and misidentification or inconsistent classification can occur due to fatigue and human subjectivity. The aim of this thesis project is to build an automated defect pattern classification system that associates to each Wafer Defect Map a class among 12 classes identified in collaboration with STMicrolectronics engineers. The 12 classes are: Normal, Donut, Incomplete, Basket Ball, Cluster Big, Cluster Small, Fingerprints, Geometric Scratch, Grid, Half-Moon, Ring, Zigzag. Normal class corresponds to normal production conditions where no pattern is present; all the other classes correspond to a real pattern and together form the Anomalous class. Our solution takes as input a WDM, determines whether a defect pattern is present and classifies it according to the set of predefined classes. In particular, we cast the automatic classification of WDMs to a image classification problem and we adopt a deep Convolutional Neural Network (CNN) to recognize the defect patterns. Practically, each wafer map is rendered to an image that is passed through the CNN classifier and gets a labeling. To treat the defect pattern classification problem as an image classification, we developed a pipeline to derive for each wafer map the corresponding image. Following this idea, we decided to exploit the well-known effectiveness and promising results of Convolutional Neural Network in image classification.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
2.73 MB
Formato
Adobe PDF
|
2.73 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/10589/140197