Schmid

Linkedin ResearchGate

Friederich Schmid

former Master/diploma student

Stay period:Dec. 10, 2020 - Jan. 10, 2022
Now: Employee at Cure Cav
Link to external profile: https://www.curevac.com/
Group: computational materials science and theoretical nanophysics

Friederich joined our chair in 2020 for his Diploma thesis done in collaboration with the Fraunhofer Institute IKTS. He defended on Dec 10, 2021, his Diploma work with the title "Automatic back end of line damage classification based on optical microscopy images utilizing machine learning methods"

Outline:
The complexity of semiconductor devices has advanced rapidly over the last years with an increasing number of features and a decrease in their size. The back end of the line (BEoL) stack has not been excluded from this development. The implementation of new materials like ultra-low-k dielectrics and decreasing dimensions of BEoL structures as well as the bumping systems have led to mechanical reliability challenges. To avoid failure in the field, the mechanical reliability must be ensured and therefore thoroughly tested already in the design phase of the product. Different strategies have been developed to obtain this. One approach is based on the induction of micromechanical load to the BEoL stack and subsequent damage analysis as depicted schematically in figure 1 and shown in [1]. The first step of the damage analysis process is imaging with optical microscopy. The current damage categorization procedure based on these images is done manually (Fig. 2), which is error-prone due to the bias of the respective evaluator and extremely time-consuming in case of larger lots which are necessary to obtain a statistical relevance for the specific experiments. Therefore, automation of the process is desired. Machine learning (ML) approaches have recently become a highly valuable and powerful tool for image processing in numerous fields [2]. Promising research has also been conducted regarding feature identification, segmentation, and categorization [3,4]. The application potential of this technology in the field of semiconductor damage analysis should be evaluated in this work.
The objective of this research is the automatic processing and classification of microscopy images of BEoL damages inflicted by micromechanical loading. To obtain this, a process flow chart should be set up and implemented in form of a Python script/software. It should take in (unprocessed) damage images of different quality (light, magnification, resolution) and process them. A toolchain should be developed to optimize the image quality for the later ML algorithm application e.g. by utilizing OpenCV functions, image segmentation, data augmentation, etc. Also, the desired damage classes should be defined based on the image feature analysis to enable the labeling of a training data set. Different machine learning models should be evaluated for the damage classification task and promising ones should be selected. The models do not have to be designed from scratch but pre-trained ones (e.g. from the touch vision package) can be utilized. They should be trained with the pre-processed images and tested with a labeled test set. The parametrization of the selected models should be optimized to maximize the percentage of correctly categorized damages. In addition to the classification, damage segmentation should be implemented to evaluate the percentage of specific characteristic areas of the entire damage (Fig. 3).
In summary, the work should result in the development of an optimized workflow including a trained ML algorithm for damage categorization and also provide additional information on individual damages by segmentation as well as derive suggestions for optimized imaging, data handling, image pre-processing, and ML model selection and parametrization.

References:

[1] J. Silomon, J. Gluch et al., Crack Identification in BEoL Stacks Using Acoustic Emission Testing and Nano X-ray Computed Tomography, IEEE International Symposium on the Physical and Failure Analysis (IPFA), 2020 .

[2] Nature Methods December 2019 issue: "Deep learning in microscopy", https://www.nature.com/collections/cfcdjceech .

[3] O. Ronneberger, P.Fischer, and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, 2015.

[4] F. Xing, Y. Xie et.al., Deep Learning in Microscopy Image Analysis: A Survey, 2018.




Schmid

Linkedin ResearchGate

Friederich Schmid

former Master/diploma student

Stay period:Dec. 10, 2020 - Jan. 10, 2022
Now: Employee at Cure Cav
Link to external profile: https://www.curevac.com/
Group: computational materials science and theoretical nanophysics

Friederich joined our chair in 2020 for his Diploma thesis done in collaboration with the Fraunhofer Institute IKTS. He defended on Dec 10, 2021, his Diploma work with the title "Automatic back end of line damage classification based on optical microscopy images utilizing machine learning methods"

Outline:
The complexity of semiconductor devices has advanced rapidly over the last years with an increasing number of features and a decrease in their size. The back end of the line (BEoL) stack has not been excluded from this development. The implementation of new materials like ultra-low-k dielectrics and decreasing dimensions of BEoL structures as well as the bumping systems have led to mechanical reliability challenges. To avoid failure in the field, the mechanical reliability must be ensured and therefore thoroughly tested already in the design phase of the product. Different strategies have been developed to obtain this. One approach is based on the induction of micromechanical load to the BEoL stack and subsequent damage analysis as depicted schematically in figure 1 and shown in [1]. The first step of the damage analysis process is imaging with optical microscopy. The current damage categorization procedure based on these images is done manually (Fig. 2), which is error-prone due to the bias of the respective evaluator and extremely time-consuming in case of larger lots which are necessary to obtain a statistical relevance for the specific experiments. Therefore, automation of the process is desired. Machine learning (ML) approaches have recently become a highly valuable and powerful tool for image processing in numerous fields [2]. Promising research has also been conducted regarding feature identification, segmentation, and categorization [3,4]. The application potential of this technology in the field of semiconductor damage analysis should be evaluated in this work.
The objective of this research is the automatic processing and classification of microscopy images of BEoL damages inflicted by micromechanical loading. To obtain this, a process flow chart should be set up and implemented in form of a Python script/software. It should take in (unprocessed) damage images of different quality (light, magnification, resolution) and process them. A toolchain should be developed to optimize the image quality for the later ML algorithm application e.g. by utilizing OpenCV functions, image segmentation, data augmentation, etc. Also, the desired damage classes should be defined based on the image feature analysis to enable the labeling of a training data set. Different machine learning models should be evaluated for the damage classification task and promising ones should be selected. The models do not have to be designed from scratch but pre-trained ones (e.g. from the touch vision package) can be utilized. They should be trained with the pre-processed images and tested with a labeled test set. The parametrization of the selected models should be optimized to maximize the percentage of correctly categorized damages. In addition to the classification, damage segmentation should be implemented to evaluate the percentage of specific characteristic areas of the entire damage (Fig. 3).
In summary, the work should result in the development of an optimized workflow including a trained ML algorithm for damage categorization and also provide additional information on individual damages by segmentation as well as derive suggestions for optimized imaging, data handling, image pre-processing, and ML model selection and parametrization.

References:

[1] J. Silomon, J. Gluch et al., Crack Identification in BEoL Stacks Using Acoustic Emission Testing and Nano X-ray Computed Tomography, IEEE International Symposium on the Physical and Failure Analysis (IPFA), 2020 .

[2] Nature Methods December 2019 issue: "Deep learning in microscopy", https://www.nature.com/collections/cfcdjceech .

[3] O. Ronneberger, P.Fischer, and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, 2015.

[4] F. Xing, Y. Xie et.al., Deep Learning in Microscopy Image Analysis: A Survey, 2018.