Abstract
The Dipmeter Advisor system represents a landmark transition of AI from academic research to commercial industrial application. Developed over four years at Schlumberger, the system automates the interpretation of borehole dipmeter data to identify geological features. This paper examines the development process, revealing that commercial success relied heavily on robust user interfaces and signal processing integration rather than just the inference engine. Key findings include the disproportionate weight of UI code compared to AI logic and the necessity of rapid prototyping in evolving expert domains.
Introduction
In the early 1980s, expert systems were primarily academic curiosities. Schlumberger sought to bridge this gap by applying AI to a high-stakes industrial problem: dipmeter interpretation. A dipmeter is a tool lowered into oil wells to measure rock conductivity in various directions, allowing geologists to determine the "dip" (tilt) of underground layers. This process is crucial for identifying hydrocarbon reservoirs, but the data is inherently noisy and requires highly skilled experts to recognize complex patterns.
The challenge for the Dipmeter Advisor was to automate this "signal-to-symbol" transformation in a way that was fast, accurate, and usable by field engineers. This required moving beyond simple rule-based systems to a comprehensive interpretation environment.
Methods and Architecture
The Dipmeter Advisor was built as a hybrid system combining traditional signal processing with heuristic AI. Its architecture featured several key components:
- Production Rules: The system utilized approximately 90 rules, partitioned by function into structural and stratigraphic categories.
- Inference Engine: A forward-chaining engine resolved conflicts based on rule order to derive geological conclusions.
- Blackboard Architecture: Conclusions were stored on a partitioned "blackboard" with 15 layers of abstraction, ranging from raw patterns to complex geological features like reefs or channels.
- Feature Detection: Custom algorithms were developed to detect "tadpole" patterns in raw data before the AI rules were applied.
Originally developed on DEC 2020/VAX hardware, the system was eventually migrated to the Xerox 1100 (Dolphin) workstation. This move was critical to take advantage of sophisticated graphical user interfaces and dedicated processing power required for real-time interpretation.
Key Insights and Results
One of the most significant findings of the project was the breakdown of the system's code. The actual AI logic—comprising the inference engine and knowledge base—accounted for only 30% of the total system. In contrast, the User Interface accounted for 42%. This highlighted that for a commercial system, usability and the support environment are as critical as the underlying "intelligence."
The system achieved substantial efficiency gains, reducing the time required to detect patterns and determine structural dip from 18 minutes to under 2 minutes. Furthermore, the development process revealed that experts are "moving targets"; as the senior expert interacted with the system, their own understanding of the domain evolved, requiring the system to serve as a testbed for new ideas.
Discussion and Limitations
Despite its success in technology transfer, the Dipmeter Advisor faced several limitations. Its handling of uncertainty remained rudimentary, often relying on simple rule ordering. Additionally, the system lacked a "global" model of geological processes, leading to a localized view of vertical consistency in rock layers. The authors concluded that such core reasoning deficiencies could not be solved by adding more rules but would require a fundamental redesign of the expert system tools to incorporate broader geological perspectives and multiple expert viewpoints.
Summary updated