Mineral Exploration Using Modern Data Mining Techniques

Returns from gold exploration have been disappointing over the last 20 years, despite the surge in quality and quantity of exploration data. Historically, major discoveries have occurred in waves following the introduction of new methods. This paper argues that the new methods driving the next wave of discoveries will be found in recent developments in data mining techniques, including visualization and probabilistic modeling. Visualization techniques present information to the brain in ways that allow patterns to stand out and be more readily perceived by our own human intelligence. Combined with geophysical inversion, these techniques make it easier to integrate multiple data sets and to build geologic models which fit current knowledge and understanding. These models can then be passed among, and visually shared by, workers from all the exploration disciplines. Probabilistic modeling techniques provide an estimate of the probability that some location with given exploration characteristics hosts a deposit, based on a set of known examples. The weights of evidence approach, which has already been used for this purpose, can provide useful results, but is limited by its basic assumptions. Neural network and kernel methods, on the other hand, are not limited in this way and can extract more meaningful information from data. The approach is demonstrated by a study of gold exploration in the Walker Lane, a mature mining district straddling the Nevada-California border in the western United States. This study incorporates 25 primary exploration data layers including geology, remote sensing, geochemistry, gravity, aeromagnetic and radiometric surveys, digital terrain and regional structure, together with known gold deposits. Care is needed in presenting data to the model. Geophysical data, for instance, may have little significance as point values, and need an encoding that represents the pattern of data in the neighborhood of a given station. The same is true of regional structure and, to some extent, of geology. The number of inputs to the model can grow in this way into the hundreds, so that efficient optimization and regularization are required. The model allows the results for individual data sets to be analyzed separately. The geology, for example, shows a strong correlation between the known gold deposits and a Tertiary andesite. The other data sets show similar but not necessarily coincident patterns. The data sets can then be combined to produce an integrated target favorability map. A subarea of the Walker Lane that falls within the Nevada Test Site illustrates the approach. Two specific targets are identified, which would certainly be followed up if this former nuclear weapons testing area was not off-limits to exploration. Finally, the distributions of favorability scores, over the known gold region and the region as a whole, determine the probability that a location scoring higher than a given threshold hosts a deposit. The distributions of scores also permit the expected costs and benefits of an exploration program to be calculated, and show how improved targeting derived from the model reduces exploration costs and increases the probability of success.