Virtual Metrology
Machine‑learning–based virtual metrology system for predicting wafer properties from process data and enabling real‑time quality control and yield prediction.
Overview
Virtual metrology (VM) aims to predict wafer properties from equipment and process data without performing expensive and time‑consuming physical measurements. During his tenure at Samsung’s AI Center, Sung‑Un Park developed and applied a machine‑learning–based virtual‑metrology system for semiconductor fabrication. The system leverages sensor data and equipment parameters to estimate wafer characteristics, enabling on‑the‑fly quality control and yield prediction in manufacturing lines cv.stoz.kr . VM reduces reliance on physical metrology tools and shortens feedback loops, thereby improving throughput and reducing manufacturing costs.
Contributions
Process‑data modelling: Designed algorithms that ingest high‑dimensional process parameters and sensor readings to predict wafer properties, allowing engineers to monitor equipment health and detect anomalies in real time.
Real‑time deployment: Integrated the VM models into the manufacturing pipeline so that predictions can be generated during production, enabling proactive adjustments and yield optimisation cv.stoz.kr .
Collaboration with metrology experts: Worked with process engineers to interpret model outputs and refine the VM algorithms for various process steps and materials.
Related Patents
Scanning electron microscope (SEM) image correction – Co‑inventor of a method that corrects SEM images taken before and after equipment maintenance. The technique acquires two images of a wafer at different times, uses a neural network to estimate a calibration factor and applies it to correct images of other wafers blog.stoz.kr . This patent enables consistent depth measurements for SEM‑based metrology.
AI model training using domain similarity – Co‑inventor of a training scheme that adapts a pre‑trained model to a new target domain through low‑rank adaptation and bias‑term fine tuning blog.stoz.kr . The method allows efficient transfer of metrology models across different equipment types or process conditions with less than 1 % parameter fine‑tuning.
These patents provide key building blocks for the VM system by ensuring stable SEM measurements and enabling rapid adaptation of machine‑learning models to new processes.