Automated machine learning pipeline for geochemical analysis

Abstract

Biplot diagrams are traditionally used for rock discrimination using geochemical data from samples. However, this approach has limitations when facing a high number of variables. Machine learning has been proposed as an alternative to analyze multivariate data for more than 70 years. However, the application of machine learning by geoscientists is still complicated since there are no tools that propose a pipeline that can be followed from preparing the data to evaluating the models. Automated machine learning aims to face this issue by automating the creation and evaluation of machine learning models. The contribution of this work is twofold. First, we propose a methodology that follows a pipeline for the application of supervised and unsupervised learning to geochemical data. Both methods were applied to a dataset of granitic rock samples from 6 blocks in the Peninsular Ranges and the Transverse Ranges Provinces in Southern California. For supervised learning, the Decision Trees model offered the best values to classify the samples from this region: accuracy: 87%; precision: 89%; recall: 89%; and F-score: 81%. For unsupervised learning, 2 components were related to pressure effects, and another 2 could be related to water effects. As a second contribution, we propose a web application that follows the proposed methodology to analyze geochemical data using automated machine learning. It allows data preparation using techniques such as imputation and upsampling, the application of supervised and unsupervised learning, and the evaluation of the models. All this without the need to program.