The Expert System Approach to Fingerprinting Past Tectonic Settings

Geochemical fingerprinting of past tectonic environments using ancient volcanic rocks is now a well-established method and a large number of empirically-derived geochemical discriminant diagrams have been published for this purpose. The best of these have potential advantages over quantitative discriminant methods because petrogenetic modelling enables the data distribution to be interpreted in terms of process as well as simply tectonic environment. However, there are also significant downsides to the empirical diagrams. Most importantly, the concept of tectonic setting has evolved from the simple three-fold classification (ridge, arc, intraplate) of the early 1970s to include a complex array of transitional settings including syn- and post-collision environments, volcanic- and non-volcanic rifted margins and oceanic plateaus. One consequence is that few compositions are unique to any one setting, so it is even more important to view classifications in terms of probabilities rather than `black-and-white' interpretations. Another is that non-geochemical information is now even more important in making a classification: for example, when dykes of MORB composition invade continental crust during ocean opening and so must have formed at a continental margin rather than ocean ridge. ESCORT (Expert System for Characterisation of Rock Types: Pearce, J.A., 1987. J. Volcanol. Geotherm. Res. 32, 51-65) provides methodologies for combining geochemical and non-geochemical probabilities. Although much more data are now available, modern, web-based volcanic databases provide easy opportunities for upgrading the system. As with other expert systems, ESCORT comprises a Knowledge Base and an Inferencing Engine. The Knowledge Base contains probability data for each of the chosen tectonically-defined magma types. It is made up of a set of dispersion matrices from which geochemical-based probabilities can be calculated from probability density functions, together with a matrix of probabilities for various non-geochemical criteria (phenocryst types, intercalated sediments, flow type etc.) and a set of a-priori probabilities. The Inferencing Engine is based on Bayes Decision Rule, adapted to take into account uncertainties in geological evidence, which enables the different probabilities to be numerically combined. The output is a set of classification probabilities for each unknown comprising an assignment probability for each tectonically-defined magma type. Upper and lower thresholds, calculated using a training set of rocks with known affinities, form the basis of the final classification. Magma types with probabilities below the lower threshold may be statistically rejected and those above the lower threshold may be regarded as possible. The magma type with the highest probability may be assigned with confidence if it also lies above the upper threshold; otherwise it may be regarded only as the `most likely' classification.