AN INTRODUCTION TO STARMAG^SM

STARMAG^SM is a neural-network based system for structure identification from magnetic data. Recently, the technique has been extended to include analysis of gravity data as well; this point will be further developed later. This document explains the way that STARMAG^SM works, compares it with other semi-automatic quantitative analysis tools, and describes some recent developments and current work on the technique.

Like most other neural network applications STARMAG^SM is based on the concept of supervised learning. This means that in order to produce useful output, the program must first be presented with a set of data generated by known structures of the type which is to be identified. The output of the program is compared with the desired response for each input, and the network is adjusted until an appropriate match between input and output data is obtained. The program is then ready to analyze data in which the location of the structure of interest is unknown.

A typical application of STARMAG^SM is the prediction of residual basement uplift. Often, shallow structural oil and gas plays are controlled by comparatively subtle basement uplifts, either by flexure or in the form of back-tilted fault blocks. If the basement is magnetic, these uplifts lead to characteristic anomalies in the magnetic field which are often considerably smaller than other anomalies on a profile but which have a characteristic shape. Given an approximate basement shape for a particular area, STARMAG^SM can be trained to recognize these characteristic anomaly shapes and to estimate the height and width of the basement uplift. This strategy has proven successful in predicting prospective areas in a number of studies over the past five years.

The neural network strategy contrasts sharply with typical potential-field depth estimation methods in that the program can be "tuned" to a particular type of target for each problem. Whereas most inverse programs have a comparatively limited repertoire of target types, STARMAG^SM can be adjusted to almost any conceivable problem. This makes the method particularly well-suited to problems where the target structure is well-understood, and the selectivity can be used to some extent to suppress the selection of features which are not of interest.

No automatic potential-field interpretation method can escape the usual problems of ambiguity. STARMAG^SM, for example, cannot easily distinguish between a small basement uplift at depth and the contact between basement rocks of different susceptibility. Apparent susceptibility mapping on basement can assist in resolving this ambiguity, but recently another tool has been developed as well: neural network gravity analysis. This requires high-quality gravity data and an appropriate geological setting, but has proven to be very effective in confirming structures delineated by STARMAG^SM.

Development of new STARMAG^SM applications continues. Areas under investigation include mining-scale applications, including the detection of kimberlites, and three-dimensional (grid-based) analysis. We anticipate a wide variety of such specialized uses for this tool will develop over the next few years.