Skip to content
Theme
OnDevice Tools

Tech BusinessChips & Semiconductors

CMP MRR prediction framework uses self-supervised learning

Researchers from SUTD, A*STAR and NUS propose GREP, an attention-based framework that learns from unlabelled sensor signals before supervised fine-tuning for CMP MRR prediction.

Chip wafer with sensor data feeding a virtual metrology model

Key takeaways

  • Semiconductor Engineering reports a paper from SUTD, A*STAR and NUS on virtual metrology for CMP MRR prediction.
  • GREP uses global proxy-target regression on unlabelled sensor signals, followed by supervised fine-tuning.
  • The evaluation compares GREP with a feature-engineered supervised baseline under limited labels, sensor noise and process drift.
  • The available excerpt does not provide numerical performance results.
In this article
  1. What the paper proposes
  2. How GREP learns
  3. Tests under difficult conditions
  4. Research context
  5. Why it matters for chip making

What the paper proposes

Semiconductor Engineering reports that researchers at SUTD, A*STAR and NUS have published a paper on virtual metrology. The target is CMP MRR prediction. The work proposes Global Regression Pretraining, or GREP. It is a self-supervised and attention-based framework for estimating MRR from sensor signals.

The paper appears as an arXiv preprint, arXiv:2611.16172, from 2026. It is listed under a CC BY-NC-ND 4.0 licence. The authors are Jonathan Adam Rico, Nagarajan Raghavan and Senthilnath Jayavelu. The paper lists three Singapore research organisations as affiliations.

CMP is the chip-making process named in the paper, and MRR is the measure the model is trained to predict. Virtual metrology refers to using data-driven models to estimate such measurements. The study focuses on how such a model can learn when labelled examples are scarce.

How GREP learns

Illustration: AI, settings, internet, market trend and the labels CMP, MRR, Nus

The GREP method uses two stages. In the first stage, it applies global proxy-target regression to sensor signals that have no labels. This allows the model to learn time-related patterns without MRR targets. The stage is self-supervised because it does not need labelled MRR examples.

The second stage is supervised fine-tuning. In this stage, labelled examples adjust the model for MRR estimation. This form of learning relies on labelled examples. The paper describes GREP as self-supervised because the first stage learns from unlabelled data.

The paper also describes GREP as attention-based. In machine learning, attention methods typically weigh different parts of an input. The available excerpt does not provide numerical scores, but it names the comparison method: a feature-engineered supervised baseline.

Tests under difficult conditions

The researchers evaluate GREP against a feature-engineered supervised baseline. They do so under three conditions: limited labelled data, sensor noise and process drift over time. These conditions change the amount of labelled data, the quality of sensor signals or the stability of the process.

Limited labelled data means fewer examples with known MRR values are available for supervised training. Sensor noise means the input signals may contain errors or random variation. Process drift over time means the process behaviour can change as time passes.

The paper does not give detailed results in the available excerpt. It does show that the evaluation is designed to test the framework in settings named by the authors. The comparison with a supervised baseline provides a reference point for assessing the approach.

Research context

The study sits within virtual metrology research for chip manufacturing. Virtual metrology models use available data to estimate process measurements. In this case, the target is MRR in CMP. The paper proposes a learning method that can be pre-trained without MRR labels.

The use of unlabelled sensor signals is central to the approach. Self-supervised learning allows a model to learn from data that lacks explicit target labels. This differs from supervised learning, which needs labelled targets. The paper then fine-tunes the model for the specific MRR prediction task.

The evaluation conditions are also important. Limited labels, noisy sensors and process drift are common concerns in data-driven manufacturing research. By naming these conditions, the paper frames the problem in practical terms.

Why it matters for chip making

Chip production depends on accurate process control. When a model can estimate a metric from sensor data, it may support monitoring and control. The paper addresses this need through CMP MRR prediction.

The self-supervised design is relevant where labelled data is limited. The evaluation also considers sensor noise and process drift. These are issues that can affect models used in manufacturing environments.

The work is a research preprint, not a commercial product announcement. Semiconductor Engineering lists it among test, measurement and analytics technical papers. Readers can consult the arXiv entry for the full method and evaluation. The paper's contribution is methodological, because it presents a framework and an evaluation rather than a commercial product at present.

Sources

  1. Self-Supervised Virtual Metrology For CMP MRR Prediction (SUTD, A*STAR, NUS)Semiconductor Engineering

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.