PROSPECTOR finds molybdenum

In Science of 3 September 1982 Campbell, Hollister, Duda and Hart reported that the expert system PROSPECTOR, given geological maps drawn before drilling and rules from a porphyry molybdenum specialist, identified on Mount Tolman in Washington State the location of previously unknown ore-grade mineralization.

The system was built in the Artificial Intelligence Laboratory of SRI International; at the time of the paper Duda and Hart were at the Fairchild Laboratory for Artificial Intelligence Research in Palo Alto, and the rules came from a porphyry molybdenum specialist. The Christian Science Monitor of 30 September 1982 quoted Rene Reboh, project director of SRI's Expert Systems Program: each expert was questioned for about fifty hours, and program and expert agreed on the same problems to within 6 or 7 percent on average; those are the developer's words. How PROSPECTOR reasoned is set out in an SRI technical note of June 1981. Every statement in a model has a prior probability; a rule multiplies the odds of a hypothesis by a sufficiency measure when the evidence is observed and by a necessity measure when it is absent, both given by the specialist. The geologist answers on a certainty scale from -5 to +5, where 0 means nothing is known; 'and' takes the smallest of the probabilities, 'or' the largest. To choose drilling sites the system took digitized maps, divided them into a grid of 128 by 128 points and ran the network 16,384 times to produce a colour-coded favourability map. The prediction is older than the paper. In BYTE of September 1981 (p. 262) Duda and Gaschnig wrote that PROSPECTOR had made its first prediction of a molybdenum deposit's location and that drilling was under way to test it; the caption of its Mount Tolman photograph says ore-grade molybdenum had been found earlier in a bright area on the right and more was predicted in an undrilled one on the left. Technology (Jan.-Feb. 1982, p. 70) calls it accurate. The Science paper, received on 20 April 1982, is dated 3 September 1982, the date this record carries; the prediction was made before the BYTE issue of September 1981, and no document read gives its day. The record can state the confirmation only in outline. The Science paper's full text was not read (its page shows only the abstract, references and dates), and the 1981 SRI note does not mention Mount Tolman. The Handbook of Artificial Intelligence (volume II, 1982) says PROSPECTOR recently predicted the location of molybdenum ore at a site in Washington State and that subsequent drilling substantially confirmed it, without naming Mount Tolman, a company, hole or date. Nilsson (2009, pp. 298-299) quotes, with ellipses, a letter of 1985 from Duda, Hart and Reboh to the journal Artificial Intelligence (the letter itself was not read): a large mining company had found an ore body by drilling over 200 holes in one region; its further drilling matched the favourability map, verifying the predicted large, previously unknown region of ore-grade mineralization and the barren areas; the area had since become unprofitable to mine. Nilsson adds that PROSPECTOR did not find a deposit at a site previously unexplored. The company is not named. Independent geological documents describe the deposit without naming PROSPECTOR. By a 1990 report of Washington State's Division of Geology and Earth Resources, Bear Creek Mining drilled 45,000 feet there in 1964-1976; AMAX Exploration signed an agreement in 1978, completed 386,827 feet of drilling in a two-year programme (the drilling alone cost $10,167,739 from 18 August 1978 to December 1980) and gave up its lease in 1982; proven and probable reserves are put at 2.4 billion tons of ore at 0.093% MoS2 and 0.09% copper. A US Bureau of Mines circular gives 900 million tonnes at 0.06% molybdenum from 91,140 m of diamond drilling, and the state's geological newsletter of January 1980 reports AMAX's drilling as under way. So the deposit as a whole was known and had been drilled since the 1960s; the paper's claim concerns mineralization not known when the maps given to the program were drawn. Size and value: the only figure for the size of the ore PROSPECTOR found, in an Eos note of 1982 on the Science paper (source not stated), is a subsurface deposit with an areal extent over 18 square kilometres, no tonnage given. That exceeds the 1.6 by 3.2 km plan of the whole orebody in a 1995 abstract; the two are not reconciled. A 1988 Air Force thesis on another subject, citing a 1985 magazine article not read, says PROSPECTOR led geologists to a $100 million deposit; no document read from 1981-82 values the ore. Later retellings, not relied on: SRI's company history (about 2004) puts the bed under smelter tailings, which no document read mentions; Peter Hart's 2018 oral history, a memoir, thinks it was in Idaho and says it was mined in a big open pit, contradicted by the 1990 report, the Bureau of Mines circular and the 1985 letter. What the record does not claim: which holes confirmed the prediction, who drilled them and when, what the ore was worth, or that the company of the 1985 letter was AMAX. The full text of the Science paper, the letter itself and SRI's final reports of 1978-80 were not read; the Eos note and the 1995 paper were read only as abstracts.