Large Pre-Trained Models and Few-Shot Fine-Tuning for Virtual Metrology: A Framework for Uncertainty-Driven Adaptive Process Control in Semiconductor Manufacturing
Chin-Yi Lin, Tzu-Liang (Bill) Tseng, Solayman Hossain Emon, Tsung-Han Tsai
AI summary
Problem
Traditional virtual metrology models degrade under frequent recipe changes and equipment shifts, while expensive physical measurements limit the labeled data available for training and adaptation.
Approach
A generative foundation model learns historical sensor data to predict wafer quality and uncertainty, triggering physical measurements only for high-uncertainty wafers, which are then used to rapidly fine-tune the model with just a few samples.
Key results
- Achieves state-of-the-art accuracy with significantly reduced metrology overhead
- Enables rapid model recalibration using only 1–5 new labeled wafers
- Provides theoretical convergence proofs and label cost bounds
- Implements dynamic uncertainty threshold adjustment for long-term stability
Why it matters
Semiconductor manufacturers and process control engineers can deploy this system to minimize measurement costs, maintain consistent product quality, and adapt to evolving production lines without costly retraining.
Abstract
High-precision wafer metrology poses significant cost and throughput challenges in modern semiconductor manufacturing, where frequent process changes and recipe variations demand highly adaptive and scalable solutions. In this paper, we present a Generative-FewShot-Active Virtual Metrology (GFA-VM) framework that unifies large-scale generative modeling, few-shot fine-tuning, and uncertainty-driven active sampling into a single, data-centric system. A foundational generative model, built on a hybrid architecture of Transformer networks and Variational Autoencoders (VAEs), learns diverse sensor characteristics in an offline stage without relying on extensive labeled data. During online inference, the model produces both wafer quality predictions and predictive uncertainties; samples exceeding a dynamic uncertainty threshold are selected for physical measurement and few-shot model recalibration. This selective sampling both reduces measurement costs and adapts rapidly to new process conditions (e.g., novel recipes or equipment upgrades), requiring only a handful of freshly labeled wafers. The paper further addresses the long-term stability of the system through a self-updating mechanism that adjusts the uncertainty threshold when distributional shifts occur. Empirical evaluations confirm that our GFA-VM approach achieves state-of-the-art accuracy while significantly reducing metrology overhead compared to conventional virtual metrology methods. Additionally, rigorous theoretical analyses—including proofs of convergence and label cost bounds—demonstrate the reliability of using a generative foundation plus meta-learning technique. By fostering on-demand adaptation within a closed- loop framework, GFA-VM offers a comprehensive, scalable strategy for next-generation semiconductor process control. Note to Practitioners—Today’s semiconductor fabrication processes involve numerous machine types and recipe variations, making frequent measurements both expensive and time- consuming. The approach described in this paper, Generative- FewShot-Active Virtual Metrology (GFA-VM), aims to help practitioners maintain consistent product quality while significantly reducing the number of physical wafer measurements. Traditional virtual metrology solutions often require extensive retraining when new recipes or equipment changes occur; our method mitigates these burdens by using a “foundation model” that learns from large volumes of historical (both labeled and unlabeled) sensor data and can be quickly adapted with only a few newly measured wafers. Practically, engineers can integrate GFA-VM into existing Manufacturing Execution Systems (MES) or Advanced Process Control (APC) platforms. The system continuously estimates each wafer’s quality and flags only those with high uncertainty for actual measurement. This targeted approach optimizes the use of metrology resources and speeds up decision-making. The key benefit is flexibility: each time a new recipe or tool is introduced, the model can be recalibrated using as few as one to five measurements. However, initial setup requires careful data gathering to train the generative model, and ongoing tuning depends on reliable sensor signals. In addition, practitioners should note that sudden, large-scale process changes still need more measurements for model stability. Looking forward, the same strategy can be applied to other manufacturing contexts—where high-dimensional sensor data and limited measurement capacities challenge real-time quality control.