EdgeAI has the power to make production processes more efficient, reliable and easily adaptable to different market needs. To guarantee a sustainable and resilient European Digital Industry we need a semi-autonomous and AI-supported production environment. By applying digitized control systems we can develop such smart and resilient industries. These advanced process control systems include advanced sensing, automated defect classification, the implementation of a digital tacit knowledge platform and virtual metrology. In this blog we will focus on the latter.

What are the challenges we face in semiconductor production?

The semiconductor industry is a highly dynamic field. The life cycle of new products is typically short, accordingly production processes need to be adapted quickly. Also, the quality standards are really high for microelectronics and strict quality criteria have to be met. These high requirements are very time and resource intensive and cause the following challenges:

  • quality control measurements are necessary after every key production step
  • different machines are needed for single production steps, complicating communication between different devices/interfaces
  • machines from different suppliers and of different generations generate non-consistent output data
  • heterogeneity of machine, measurement and logistic data

Virtual metrology can be used as a tool for process control and quality monitoring. Here, we want to show in detail the potential and advantages of virtual metrology as a tool in semiconductor manufacturing. In the course of this project a demonstrator will be developed to allow for smart and efficient controlling within the semiconductor production and processing.

Virtual metrology aims to fulfill the following three goals:

  1. Quick identification of deviations from standard production results.
  2. Reduction of production cost and time by minimizing the physical steps during the flow of production.
  3. Supporting the continuous improvement of new process steps.

In this project, multiple data sources have to be transferred for virtual metrology from the Real-Time (RT) systems to a so-called ‘Data Handling Layer’. These RT systems (i.e. the production equipment, the material tracking system and advanced process control systems) are mostly commercial solutions, where the system architecture is given by the vendor with limited degree of freedom for specific requirements of the output data. Furthermore, they mostly rely on common database management systems and do not feature tailored options for specific requirements. From the consolidated Data Handling Layer, data can be generated that serves as input data for an AI model. This data can be used both for model training and for predictive purposes within the production process. In the latter case, the prediction result is then fed back into the Data Handling Layer, where it is available for end-systems, e.g. the manufacturing execution system or other applications such as statistical process control (SPC) systems.

 

The Edge of Advantage: Why do we need Virtual Metrology?

The new technology offers several advantages for the production and processing of semiconductors.

  • Standardization of data in the production via the established Data Handling Layer
  • reduced number of physical measurements
  • no limit of measurement sampling frequency with virtual metrology
  • real-time feedback to the operator
  • live and online status of every RT system
  • predictive power for next production step and next required production equipment
  • savings in time and costs by higher efficiency in production

The role of ams OSRAM and project status

ams OSRAM helps in developing virtual metrology for semiconductor production. At the current status of the project the preparations and setup of the Data Handling Layer are finished. As a first step, the algorithm for handling the different types of data during production had to be chosen. After an extensive evaluation of different algorithms, we have decided to use Gradient-Boosting Trees as a prediction algorithm. Gradient Boosting Trees is an ensemble learning method, which is attractive due to its handling of missing data, its resistance to overfitting, and its suitability for large tabular datasets, as is the case of our AI use cases.

As a second step, we needed to choose a software to allow for machine-learning of the AI model in virtual metrology with the delivered data.  The software framework, which was finally chosen to train the AI models, was Catboost, a machine-learning library designed explicitly for handling tabular data effectively. Catboost is well-known for smoothly performing regression and classification tasks for tabular data and was the perfect fit for the task. Furthermore, its simplicity as a tree-based library makes it suitable for deployment in different types of computational platforms without requiring specific hardware components.

With these preparations we implemented our Data Handling Layer, a consolidated data source with a completeness of >99%. We thereby can generate defined datasets with attributive, sensor and measurement information for training and prediction from several sources. For the next steps we have to evaluate data and model stability, do a reliability analysis and further conduct the strategy development. For the further outlook it is planned to verify, validate and test the demonstrator for virtual metrology.

Blog signed by: ams-OSRAM team

Share This