Chips & deep tech · Article
Researchers Propose GREP Method to Predict Chip Polishing Removal Rates Without Many Labels
Semiconductor Engineering ·

At a glance
Researchers from SUTD, A*STAR, and NUS proposed a method that predicts how fast a chip-polishing machine removes material, using a model trained mostly on unlabeled sensor data. The method, called GREP, first learns patterns from the sensor signals and is then fine-tuned with a smaller set of labeled measurements. The authors tested it against a conventional approach under limited labeled data, sensor noise, and changes in the process over time.
Key points
- The model is trained first on raw sensor signals without labels, then fine-tuned on labeled data.
- The removal rate is the speed at which the polishing process removes material from a wafer.
- It was compared with a standard supervised method under three harder conditions.
- The conditions were few labeled examples, noisy sensors, and drift in the process over time.
Why it matters
Chip polishing is a key manufacturing step, and predicting its results without costly measurements could make production more efficient.
Original source
Self-Supervised Virtual Metrology For CMP MRR Prediction (SUTD, A*STAR, NUS)
Semiconductor Engineering · Nov 17, 09:56 KST
Read on Semiconductor Engineering ↗AI summarized the original article. Check the original for accuracy. Copyright belongs to Semiconductor Engineering.
More from this field
- Contract Cleanroom Workers at Samsung Chip Plant Report Nosebleeds, Rashes and Low Pay
A contract worker named Kim Jin-gi, a pseudonym, cleans and moves wafer boxes for about 12 hours a day in a Samsung Electronics chip plant in Hwaseong. He describes frequent nosebleeds, skin infections from hot, sweaty protective suits, night shifts, and fears about chemicals he has never been told about. He says workers are reluctant to report injuries because of a bonus rule, and that one-year contracts keep their pay near the minimum wage.
프레시안 · Nov 17, 08:02 KST - Blackstone's BXDC Takes 75% Stake in AWS Data Center Ventures in $5 Billion Deal
Blackstone's newly listed data center fund BXDC reportedly bought a 75% stake in two data center joint ventures run with Amazon Web Services, valued at about $5 billion. The fund will own 13 newly built powered shells, meaning buildings with power infrastructure only, and AWS will lease the whole portfolio. The deal is expected to close within the year.
뉴스핌 · Nov 17, 07:40 KST - NVIDIA Releases cuPotential to Speed Up Molecular Dynamics Simulations
NVIDIA released cuPotential, a library that speeds up molecular dynamics simulations, which are used to model how atoms and molecules move in physics, chemistry, biology, and materials science. It offers modular parts that can be plugged into existing tools.
NVIDIA Technical blog · Nov 17, 07:30 KST - TSMC Kumamoto Plant 1 Posts NT$2.4 Billion Nine-Month Profit as Plant 2 Build Advances
TSMC's first Kumamoto plant in Japan posted a net profit of 2.4 billion New Taiwan dollars over nine months. Construction of its second plant is also reported to be on track.
글로벌이코노믹 · Nov 17, 05:35 KST