How AI Found a Major Washington Mineral Deposit

In the late 1970s and early 1980s, a pioneering artificial intelligence program named PROSPECTOR made history by predicting the location of a previously unknown, multimillion-dollar molybdenum deposit at Mount Tolman in Washington State. Developed by SRI International, this early expert system translated the complex knowledge of human geologists into mathematical inference rules. By processing surface data, fault lines, and rock compositions through a network of probabilities, PROSPECTOR successfully pinpointed where exploratory drilling should occur, proving for the first time that AI could solve real-world physical exploration challenges.

The Architecture of PROSPECTOR

PROSPECTOR was designed to replicate the decision-making process of master geologists. Human experts studied specific types of ore deposits and helped engineers translate their intuition and field knowledge into a machine-readable format. The system did not rely on modern machine learning; instead, it used a hardcoded knowledge base consisting of more than 1,000 IF-THEN rules.

These rules formed an "inference network," which connected field observations (the evidence) to geological conclusions (the hypotheses). For example, an observation might be the presence of a specific altered mineral like quartz or sericite. The system would link this observation to a higher-level hypothesis, such as the likelihood of hydrothermal activity necessary to form a porphyry molybdenum deposit.

Bayesian Inference and Fuzzy Logic

Geology is an inexact science, as rock data from the field is often incomplete, ambiguous, or uncertain. Simple true-or-false programming could not handle this ambiguity. To overcome this, PROSPECTOR’s inference engine relied on two mathematical concepts: Bayesian probability and fuzzy logic.

Bayesian Probability: Every rule in the system was assigned two statistical weights based on interviews with human experts. The first weight measured how strongly the presence of evidence supported a hypothesis. The second measured how strongly the absence of that evidence refuted the hypothesis. When field data was entered, the system used Bayes' theorem to update the probability of an ore deposit existing at that location.
Fuzzy Logic: Field geologists rarely say a mineral is simply "present" or "absent." Instead, they estimate its abundance. PROSPECTOR allowed users to input confidence scores on a scale from -5 (definitely not present) to +5 (definitely present). Fuzzy logic rules allowed the AI to combine these varying degrees of certainty to calculate an overall likelihood for the final conclusion.

Evaluating the Mount Tolman Data

In 1978, exploration was underway at Mount Tolman in Ferry County, Washington. Geologists had already mapped the surface, analyzed geochemical samples, and conducted preliminary drilling, yet the primary zone of the potential molybdenum deposit remained elusive.

Researchers fed the geographical and geological data from the unexplored areas of Mount Tolman into PROSPECTOR. Instead of issuing a single verdict for the entire mountain, the system broke the terrain down into a spatial grid. For each cell in the grid, PROSPECTOR ran the local geological evidence through its porphyry molybdenum model.

The inference engine propagated the field probabilities upward through its network of rules. It evaluated tectonic settings, the structure of local fault zones, rock types, and mineralization patterns. Once the calculations were complete, the system generated a digital map highlighting the exact areas with the highest mathematical probability of containing high-grade molybdenum.

The Discovery Validation

PROSPECTOR identified a specific, previously untested patch of land at Mount Tolman as having a remarkably high probability of harboring the target ore. Subsequent core drilling by mining companies at the exact coordinates suggested by the AI confirmed the prediction. The drills intersected a massive porphyry molybdenum deposit valued at approximately $100 million at the time.

This success validated the use of probabilistic inference networks in the physical sciences. By structuring human expertise into weighted rules and applying Bayesian mathematics to uncertain data, PROSPECTOR demonstrated that artificial intelligence could accurately deduce what lay beneath the Earth's surface.