CORE's technical results in the iQonic project
iQonic project
CORE's technical results
Author: Spiros Fourakis
22nd September 2022
Figure 1: Original wafer image (left) and corresponding defect detection result from CORE defect detection model (right).
The iQonic project is entering its last few months and the project’s final webinar will take place in the morning of September 22nd. The project centers around a scalable zero-defect manufacturing platform covering the overall process chain of optoelectrical parts, facing the challenge of dealing with the evolution of the equipment, instrumentation and manufacturing processes they support. CORE’s efforts focused on deep learning algorithms to ensure strong prediction and detection skills and respective reactions to achieve zero defects.
More specifically, within the iQonic project, CORE has developed a new complete framework for defect detection and quality prediction of final assembled product in two demo cases AlPES and Prima. In particular, CORE’s contribution for Alpes Demo Case concerns the development of a machine learning-based defect detection solution which is focused on defect identification on the wafer parts (Figure 1).
Figure 2: Validation in early anomaly detection
Concerning the Prima Demo Case, a new and complete framework for prediction quality of a multi-laser emitter product, based on deep learning models was developed. The framework consisted of two stages: (1) early anomaly detection, focused on investigating the suitability of the final assembled product during early production stages, and (2) accurate prediction, which focused on estimating the quality index of the final product from its’ early production stage. Both models were successfully validated with real offline data from the production line. Especially, the anomaly detection model correctly predicted all the normal assemblies and nearly all defective assemblies, with only 3 false negatives (Figure 2).
Figure 3: Validation in accurate prediction of total power
Similarly, the quality prediction model demonstrated considerably low prediction errors and good generalisation performance (Figure 3).
