24-10 Sediment source distribution and mixing identification with multitype provenance dataset via tensor factorization

Session: Advances and Applications in Geochronology for Interpreting Stratigraphic and Basin Records, Part I

Presenting Author:

Authors:

Abstract:

Modern provenance datasets are usually comprised of several methods including petrography, grain-size, U-Pb dating, whole-rock, and mineral geochemistry. Because a common goal of sediment provenance research is to identify the characteristics of sediment sources, a desirable output of a statistical analysis applied to provenance data is to estimate the composition of endmember sources of a particular group of sinks.

We propose the use of Tucker-1 tensor factorization applied to a modified petrographic data cube (PDC, Weltje, 2004) to obtain the sources and mixing proportion of sources using a petrographic (BP)[j1] [BG2]  and detrital zircon U-Pb dating and geochemistry (DZ) data. BP data were collected as paired mineral and mineral size data and so BP data are converted into KDEs representing distributions of their respective grain sizes, whereas zircon U-Pb data are presented as KDEs of age. The KDEs are discretized along 2^n+1 nodes which are evenly spaced across the distribution space. Discretized DZ and BP KDEs are then combined into a 3-way tensor (Y_ijk, i=sinks, j=features, k=distributions). To find a matrix with the sources’ proportions (A) and a tensor with the sources’ distributions (B) the tensor factorization optimizes a mixture model (Y=A*B + E), where E is the misfit between modelled tensor Y (A*B) and empirical tensor Y. The best number of sources (K<<i) is obtained by calculating the misfit for a wide range of ranks and calculating the maximum curvature of the misfit versus rank curve. We apply the model to two river catchments in Chile: El Salvador and Los Bayas. The El Salvador system is best explained by two end-member sources. The model captures mixing at tributary junctions and downstream dominance by one of the sources. In the Los Bayas system, two sources are also identified. The upstream samples are dominated by finer-grained material, whereas downstream samples show a shift toward coarser input. In both cases, grain-size variation is the main feature leveraged by the model. Analogous to the PDC, the tensor factorization operates in the granulometric realm, which allows the identification of transport invariant subpopulations. We conclude that the Tucker-1 factorization is a viable method of analysing the wealth of data available in a PDC. The method allows us to synoptically examine multiple provenance proxies, thereby providing a framework for multi-type provenance analysis.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.