Human Interpretable Virtual Metrology in the Semiconductor Manufacturing

Amina MevicAAAI 2025 · DOI

Abstract

In semiconductor manufacturing, physical inspection of every wafer is often impractical due to time constraints and the destructive nature of certain tests. Virtual Metrology (VM) offers a non-invasive alternative by predicting product quality from sensor data. This paper introduces a model-agnostic framework for VM that predicts multiple product properties simultaneously. By leveraging projective operators, the framework provides human-interpretable explanations for complex multi-output predictions faster than industry standards like SHAP, ensuring that AI-driven quality control is both accurate and transparent for domain experts.

Introduction

Semiconductor manufacturing relies on high-volume production where traditional random sampling for quality control can lead to undetected defects. Virtual Metrology (VM) addresses this by using sensor data—such as temperature, pressure, and voltage—collected during the manufacturing process to "predict" the quality of a wafer without physical measurement.

The Virtual Metrology (VM) pipeline: mapping sensor data from manufacturing equipment to multi-point wafer quality measurements.

Despite its potential, VM faces two primary hurdles: the complexity of predicting interrelated properties across multiple points on a wafer (e.g., a 17-point map) and the "black box" nature of machine learning models. Engineers require interpretable models to trust and act upon AI-driven flags. This research proposes a framework that utilizes projective operator theory to provide fast, stable, and multi-output explanations.

Methodology

The proposed framework is centered on a three-pillar approach to enhance both performance and interpretability:

  • Multi-Output Learning: Instead of treating each measurement point as an independent task, the system predicts properties across 17 points on a wafer simultaneously. This approach captures the inherent spatial correlations between different regions of the wafer, improving overall prediction accuracy.
  • Projective Selection (ProjSe): To handle the high dimensionality of sensor data, the ProjSe algorithm is employed for feature selection. Using the algebra of projection operators, it identifies the most critical sensor signals from a pool of redundant data.
  • Model-Agnostic Explanations: The core contribution is a new explanatory method grounded in projective operator theory. Unlike traditional single-output explainers, this method is specifically designed for vector-valued (multi-output) problems, providing insights into how input features affect the entire wafer map.
  • Stability Index: To ensure reliability, a "stability index" was introduced to measure the consistency of explanations across different runs or minor data perturbations.

Experimental Setup and Datasets

The research was validated using proprietary industrial data from a Physical Vapor Deposition (PVD) process, supported by Infineon Technologies. The dataset includes various sensor signals such as deposition time, power, voltage, electrical current, temperature, and pressure. Additionally, the explanatory method was tested on real-world benchmarks to demonstrate its model-agnostic capabilities across different architectures.

Key Results and Discussion

The experimental results highlight several advantages of the projective operator approach:

  1. Computational Efficiency: The proposed explanatory method significantly outperforms SHAP and TreeInterpreter in terms of computation time, making it more suitable for real-time manufacturing environments.
  2. Reliability and Stability: Despite the increased speed, the method achieves a stability index and correlation comparable to state-of-the-art interpretability tools.
  3. Physics-Alignment: Feature rankings produced by the ProjSe algorithm were found to align with the underlying physics of the PVD process. This alignment with expert intuition is critical for the adoption of AI tools in industrial settings.
  4. Inter-output Correlation: The study confirms that exploiting the "joint information" between different measurement points on a wafer enhances prediction performance compared to single-output models.

Conclusion and Future Work

This research demonstrates that projective operators provide a robust foundation for interpretable Virtual Metrology. While the current ProjSe algorithm excels at handling approximately 20 outputs (matching standard wafer maps), future work will focus on improving its scalability for lower-dimensional output spaces. Additionally, qualitative user studies with industry experts are planned to further validate the practical utility of these explanations on the factory floor.

Summary updated

BibTeX
@inproceedings{mevic2025human,
  author = {Amina Mevic},
  title = {{Human Interpretable Virtual Metrology in the Semiconductor Manufacturing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35219},
}
Figure