PIES: An Engineers Do-lt-Yourself Knowledge System for Interpretation of Parametric Test Data

Jeff Yung-Choa Pan, Jay M. TenenbaumAAAI 1986

Introduction

In the semiconductor industry of the mid-1980s, fabrication processes had evolved into highly complex sequences involving over 100 chemical and physical steps. As these processes scaled, the volume of "parametric test" data generated at the end of the production line grew exponentially. When a wafer lot failed to meet specifications, human failure analysis engineers were tasked with diagnosing the root cause from hundreds of individual measurements—a process that was increasingly slow, error-prone, and difficult to scale.

Overview of the semiconductor fabrication line and the role of failure analysis.

Traditional expert systems of the era, such as MYCIN, relied on "shallow" if-then rules. While effective in static domains, these systems struggled in the rapidly changing environment of semiconductor manufacturing. Every process update required an AI specialist to interview a domain engineer and manually update the rule base, creating a significant "knowledge engineering bottleneck."

The "Semi-Deep" Architecture

To address these challenges, the authors developed PIES (Parametric Interpretation Expert System). Unlike shallow systems, PIES utilizes a "semi-deep" architecture that organizes diagnostic knowledge into a four-level causal chain. This structure mirrors the natural reasoning process of human engineers, who think in terms of physical cause-and-effect rather than isolated logical rules.

The causal hierarchy used by PIES to link process variations to measurement deviations.

The four levels of the PIES hierarchy include:

  1. Measurement Deviations: Observed anomalies in test data (e.g., "Beta value is too low").
  2. Physical Structure Abnormalities: Deviations in the physical characteristics of the wafer (e.g., "The base layer is too thick").
  3. Process Variations: Deviations in the manufacturing steps (e.g., "The oven temperature was too high").
  4. Root Causes: The ultimate source of the failure (e.g., "A timer malfunctioned").

Knowledge Engineering for Domain Experts

A core innovation of PIES is the "Knowledge Editor," a tool designed to be used directly by semiconductor engineers without the intervention of an AI specialist. By allowing engineers to input failure cases and link them across the four causal levels using qualitative strengths (such as must, very-likely, or likely), the system shifts the responsibility of knowledge maintenance to the people on the factory floor.

The diagnostic reasoner then uses a multi-step process—hypothesization, implication, confirmation, and thresholding—to navigate this causal network. This level-by-level reasoning allows the system to quickly prune the search space, making it computationally efficient compared to flat search methods.

Evaluation and Results

PIES was rigorously tested using real-world data from Fairchild’s ISO-Z bipolar process facility in Puyallup, Washington. The evaluation involved two phases:

  • Tuning Phase: Using a set of 25 historical "problem lots," the system initially solved only 10 cases correctly. However, after engineers used the Knowledge Editor to fill gaps in the causal network, the system achieved 100% accuracy on this set.
  • Blind Testing: On 18 randomly selected new cases from production history, PIES achieved a 67% (12/18) satisfactory result rate.

Notably, in several instances, the system provided more comprehensive explanations than the original human diagnoses, demonstrating its potential to augment human expertise.

Conclusion and Future Directions

PIES represents a significant step toward making expert systems maintainable by domain experts. By decoupling the diagnostic "shell" from the domain-specific content, the authors created a tool that could capture and "package" the experience of veteran engineers for use across different fabrication sites.

While PIES still relies on experiential knowledge, its success paved the way for "Deep-Level" approaches. Future work in this area moved toward Model-Based Reasoning, where systems derive failure causes from first principles of solid-state physics and process technology rather than pre-defined causal links.

Summary updated

BibTeX
@inproceedings{pan1986pies,
  author = {Jeff Yung-Choa Pan and Jay M. Tenenbaum},
  title = {{PIES: An Engineers Do-lt-Yourself Knowledge System for Interpretation of Parametric Test Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure