In the high-stakes world of semiconductor manufacturing, precision is everything. One of the most critical steps is Chemical Mechanical Planarization (CMP)—essentially a high-tech polishing process that ensures silicon wafers are perfectly flat. To keep this process under control, engineers monitor the Material Removal Rate (MRR).
The catch? Measuring MRR physically is expensive and slow. Virtual Metrology (VM) aims to solve this by using AI to predict MRR from sensor data in real-time. However, traditional VM models often require massive amounts of labeled data or tedious manual feature engineering. A new paper by Rico et al. proposes a smarter way forward: Global Regression Pretraining (GREP).
TL;DR
The authors introduce GREP, a self-supervised framework that learns to predict chip manufacturing quality from raw sensor data without needing expensive labels upfront. By using an attention-based Transformer to "pre-train" on unlabeled signals, the model achieves a 17% improvement in handling process drift compared to traditional feature-engineered methods.
What Problem Does This Paper Address?
In a modern fab, sensors are everywhere, churning out time-series data on pressure, flow rates, and motor speeds. While this raw data is "free," the corresponding MRR labels (the ground truth) are "expensive" because they require physical measurement.
Existing VM approaches usually fall into two camps:
- Linear models that are too simple to catch complex equipment interactions.
- Feature-engineered models that rely on humans to pick the "right" statistics (like mean or variance) from the data, potentially throwing away subtle temporal patterns.
Furthermore, manufacturing equipment "drifts" over time as parts wear down, causing models trained on old data to lose accuracy. The authors sought a model that is label-efficient (works with less data), robust to noise, and resilient to process drift.
Methods
The core innovation is the GREP framework, which uses a two-stage approach:
- Self-Supervised Pretraining: Instead of starting with MRR labels, the model looks at unlabeled sensor signals. It divides each signal into three "views" (the whole sequence, the first half, and the second half). It then tries to predict "proxy targets"—mathematical summaries created by projecting these views onto random Gaussian vectors. This forces the model to learn the underlying "shape" and temporal structure of the sensor data.
- Supervised Fine-Tuning: Once the model understands the sensor signals, it is "fine-tuned" on a small set of labeled data to perform the actual MRR prediction.
The architecture relies on PatchTST, a channel-independent Transformer encoder that treats time-series data as "patches" (similar to how Vision Transformers treat image pixels), allowing it to capture long-term dependencies efficiently.
Datasets
The study uses the PHM 2016 CMP dataset, a well-known benchmark in the field. It contains 19 process variables (like slurry flow and table speed) across two different operating modes:
- Low-speed polishing (Chambers 4-6)
- High-speed polishing (Chambers 1-3)
Key Insights and Ideas
- Random Projections as "Teachers": The most surprising insight is that regressing on random projections of data is an effective self-supervised task. It doesn't require the complex "contrastive" setups (comparing similar vs. different samples) often seen in modern AI, making it computationally lighter.
- Temporal Awareness: By forcing the model to predict targets for the halves of a sequence during pretraining, the model naturally learns how a polishing run evolves from start to finish.
- Representation vs. Engineering: The results suggest that deep learning can "discover" better features for handling equipment aging (drift) than human engineers can by just calculating summary statistics.
Key Results
The authors compared GREP against a strong baseline that used hand-crafted features and tree-based regression.
- Label Efficiency: GREP remained competitive even when only 10% of the training labels were available.
- Noise Robustness: Both GREP and the baseline handled synthetic sensor noise well, showing that the attention mechanism is naturally good at filtering out "jitter."
- Process Drift (The Big Win): When tested on data from much later in the equipment's lifecycle, GREP significantly outperformed the baseline. In the final temporal interval, GREP achieved a 9.32 RMSE, a 17.0% reduction in error compared to the baseline's 11.23.
Limitations and Future Work
While promising, the authors note a few areas for improvement:
- Real-world Integration: The study focuses on prediction accuracy but hasn't yet integrated the model into a live Advanced Process Control (APC) loop to see how it affects actual wafer yield.
- Online Adaptation: The current model is static after training. Future work will look at "online adaptation," where the model continues to learn and update itself as new data rolls in, further combating process drift.
- Uncertainty: The authors suggest exploring "uncertainty-aware" sampling—essentially having the AI flag when it's unsure of a prediction so a human can step in for a physical measurement.
Summary updated