Discovering geochemical patterns using self-organizing neural networks: a new perspective for sedimentary provenance analysis

Abstract

An unsupervised neural network technique, Growing Cell Structures (GCS), was used to visualize geochemical differences between sandstones of four different sedimentary provenance groups: P1 (mafic), P2 (intermediate), P3 (felsic), and P4 (recycled). Multidimensional data of four sandstone data sets comprising major elements, log-normalized major elements, trace elements, and high field strength elements (HFSE) were projected into colored two-dimensional maps that can be visually and quantitatively interpreted. The cluster structure and variable distributions produced show that each sedimentary provenance group can be distinguished in the neural maps according to a unique combination of major or trace element concentrations. In these terms, the distinguishing features of each provenance group are: P1—high Fe 2O 3t, TiO 2, MgO, MnO, CaO, P 2O 5, Sc, V, Cr, and Cu; P2—intermediate Fe 2O 3t, TiO 2, MgO, MnO, CaO, Sc, V, and Cu; P3—intermediate to high K 2O, intermediate SiO 2 and Al 2O 3, low Fe 2O 3t and TiO 2, and intermediate to low Nb, Rb, and Th; P4—high SiO 2, Y, Nb, Rb, Th, Ba, and Zr, coupled with low Al 2O 3, CaO, Na 2O, Fe 2O 3t, MgO, MnO, and TiO 2. The elemental associations in P1, P2, and P3 reflect petrogenetic evolution of first-cycle sources, whereas the associations in P4 are compatible with the combined effects of recycling, weathering, and heavy mineral concentration.