A neural network application to assess magma diversity in the Etendeka igneous province, Namibia

Abstract

Abstract The geochemical discrimination of different magma types in Large Igneous Provinces is conventionally based on a few, pre-selected variables that are regarded to have petrological meaning. An alternative approach explored in this study is to apply the neural network technique of self-organising maps (SOM) to identify inherent groupings in data without knowledge or assumptions (unsupervised learning). The dataset used in this study comprises whole-rock analyses from extrusive (lava) and intrusive (dykes, sills) mafic suites in the Etendeka province, Namibia, taken from published sources and augmented by 103 new chemical analyses of dykes. Six SOM-classified groups are identified, which are unevenly distributed among the extrusive and the intrusive rock suites. The lava samples are dominated by just three of the six SOM groups (95% of all samples) and one group is absent entirely, whereas all six groups are present in the intrusive suite and five of them each comprise more than 5% of the samples. The geographic distribution of SOM-grouped dykes is heterogeneous and groups that are under-represented in the lava suite occur preferentially in a region of the pre-Etendeka basement where few lavas are preserved. Thus, the difference in magma diversity between intrusive and extrusive suites may be partly an artefact of erosion, which implies that a proper assessment of magma diversity in this and other LIPs must include the intrusive components. The correspondence of our SOM groupings with magma types in the Etendeka province that were established from petrologically defined variables is reasonably good for most trace-element abundances and ratios. However, some of the SOM groups have a wide range of initial Sr–Nd isotope ratios and a poor correspondence with the established magma types. We conclude that the SOM approach is useful for sorting out large and complex geochemical datasets but the method gives all input variables equal weight, which may be problematic if they have different responses to processes in the system under study (e.g., partial melting, fractional crystallisation, degassing, alteration). It is no substitute for expert petrological knowledge in discriminating genetically distinct magma types in an application like the present one.