Oil-Well Data Interpretation Using Expert System and Pattern Recognition Technique

Alain Bonnet, Claude Dahan

Introduction

In the early 1980s, researchers at Schlumberger developed LITHO, a pioneering hybrid AI system designed to bridge the gap between raw signal processing and high-level geological reasoning. LITHO integrates expert system production rules with pattern recognition algorithms to interpret "logs"—complex curves representing the physical properties of the earth measured during oil-well drilling.

The Challenge of Lithofacies Identification

In petroleum exploration, geologists analyze continuous measurements of properties such as density, electrical resistivity, and radioactivity to determine subsurface rock types (lithofacies). However, mapping these physical measurements to specific rocks is rarely straightforward. For instance, cherts and quartzites often produce nearly identical log responses.

To resolve these ambiguities, human experts rely on "external data," including geographic location, stratigraphy (the age of rock layers), and paleontology (fossil content). Prior to LITHO, computer-aided interpretation was primarily statistical and lacked the ability to incorporate these critical contextual nuances. The authors aimed to create a system capable of emulating a geologist's reasoning by combining raw sensor data with broader geological context.

Methodology: A Hybrid Architecture

LITHO employs a dual-layered architecture that merges symbolic reasoning with numerical analysis.

Symbolic Reasoning via Expert System

The system's core is built on the EMYCIN framework, utilizing approximately 500 production rules. These rules incorporate "certainty factors," enabling the system to make logical inferences even when input data is weak, noisy, or contradictory. This allows the system to weigh evidence from multiple sources to reach a robust conclusion.

Numerical Analysis and Pattern Recognition

To process the raw log curves, LITHO uses a two-step pattern recognition approach:

  1. Polygonal Approximation: A "split-and-merge" algorithm simplifies complex, noisy log curves into a series of straight-line segments. Interestingly, the authors found that allowing discontinuities between segments was beneficial, as these "ruptures" often signify sudden geological changes or bed boundaries.
  2. Syntactic Parsing: These segments are treated as "words" in a formal grammar. A finite automaton parses these sequences to identify geologically significant features, such as "megaramps" (progressive changes) or "plateaus."
Polygonal approximation of log curves for pattern recognition.

The system then executes an allocation procedure in a multi-dimensional log value space, weighted by the expert system's plausibility scores, to assign the most likely rock type to each depth interval.

Key Results and Insights

The authors evaluated LITHO using real-world data from ten wells across diverse regions, including Qatar, Syria, Greece, and Germany.

  • Expert-Level Accuracy: The lithological columns generated by LITHO were found to be remarkably similar to those manually interpreted by experienced geologists.
  • Contextual Integration: A central tenet of the system is that rock descriptions lack significance without context. For example, if the stratigraphic age is identified as "Cambrian," the system automatically expects highly compacted rock formations.
  • Robustness: By leveraging EMYCIN’s certainty factors, LITHO proved resilient against measurement errors common in harsh drilling environments. It excels at synthesizing multiple "weak" conclusions into a single "strong" geological interpretation.
A lithological column generated by the LITHO system, showing identified rock types and depths.

Conclusion and Future Directions

LITHO successfully demonstrated that a hybrid approach—combining the strengths of symbolic AI and numerical signal processing—could automate complex geological interpretation. While the initial system focused on labeling rock types, the authors noted that future iterations should aim to determine the "genesis" of rocks (e.g., identifying specific depositional environments like "mouth bar sandstones"). Additionally, the system's ability to explain its reasoning via the WHY command highlighted its potential as an educational tool for training new geologists in log interpretation.

Summary updated

BibTeX
@inproceedings{bonnet1983oil,
  author = {Alain Bonnet and Claude Dahan},
  title = {{Oil-Well Data Interpretation Using Expert System and Pattern Recognition Technique}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure