The aim of this work is to develop Deep Learning strategies to extend previous research (gentner2021dbam,gentner2020enhancing) concerning a task related to Virtual Metrology field in a semiconductor manufacturing. This intends to compare the previously developed model and other models more suited to the time series data under examination. The strategy used in this problem is based on Domain Adaptation through an adversarial training setup; the approach is recent and promising, allowing for better management in the absence of large amounts of data and making previously trained models reusable in similar scenarios. This research is encouraged for its effects in this field; further improving the performance of the models used means savings on production costs, since it would limit conformity tests which lead to product damage. This also leads to a possible saving of materials, an important aspect from an ecological point of view. Finally, it is increasingly necessary to consider models that use less data or are more suitable for similar processes, as in the case of Domain Adaptation techniques. The latter aspect is relevant because the data is expensive to obtain in the considered application.
Lo scopo di questo lavoro è sviluppare delle strategie di Deep Learning per estendere la ricerca precedente (gentner2021dbam,gentner2020enhancing) riguardante una mansione relativa all'ambito di Virtual Metrology in una produzione di semiconduttori. Questo vuole essere un confronto tra il modello sviluppato in precedenza e alcuni modelli più adatti ai dati di serie temporali in esame. La strategia utilizzata in questo problema si basa sul Domain Adaptation attraverso un'impostazione di Adversarial Training, un approccio recente che permette una miglior gestione in assenza di grandi quantità di dati e rende i modelli precedentemente addestrati riutilizzabili in scenari simili. Questa ricerca è incoraggiata per gli effetti che può portare in questo campo. Migliorare ulteriormente le prestazioni dei modelli utilizzati significa risparmiare sui costi di produzione, poiché permetterebbe di limitare ulteriormente test di conformità che portano al danneggiamento dei prodotti. Ciò comporta anche un possibile risparmio di materiali, aspetto importante dal punto di vista ecologico. Infine, è sempre più necessario considerare l'uso di modelli che utilizzano meno dati o sono più adatti a processi simili, come nel caso delle tecniche di Domain Adaptation. Questo è importante perché, come nell'applicazione considerata, i dati sono costosi da ottenere.
A Domain Adaptation approach for sequence modeling through Deep Learning in semiconductor manufacturing: adversarial training setup with Temporal Convolutional Network and Long-Short Term Memory models
DALLA ZUANNA, FILIPPO
2021/2022
Abstract
The aim of this work is to develop Deep Learning strategies to extend previous research (gentner2021dbam,gentner2020enhancing) concerning a task related to Virtual Metrology field in a semiconductor manufacturing. This intends to compare the previously developed model and other models more suited to the time series data under examination. The strategy used in this problem is based on Domain Adaptation through an adversarial training setup; the approach is recent and promising, allowing for better management in the absence of large amounts of data and making previously trained models reusable in similar scenarios. This research is encouraged for its effects in this field; further improving the performance of the models used means savings on production costs, since it would limit conformity tests which lead to product damage. This also leads to a possible saving of materials, an important aspect from an ecological point of view. Finally, it is increasingly necessary to consider models that use less data or are more suitable for similar processes, as in the case of Domain Adaptation techniques. The latter aspect is relevant because the data is expensive to obtain in the considered application.File | Dimensione | Formato | |
---|---|---|---|
DallaZuanna_Filippo.pdf
accesso aperto
Dimensione
2.94 MB
Formato
Adobe PDF
|
2.94 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/30823