Self-supervised Automated Mineralogical and Chemical Analysis for Hyperspectral Datasets

Identification of unknown micro- and nano-sized mineral phases is commonly achieved by analysing chemical maps generated from hyperspectral datasets, particularly scanning electron microscope - energy dispersive X-ray spectroscopy (SEM-EDX). However, the accuracy and reliability are limited by subjective human interpretation and instrumental artefacts in the chemical maps. At the same time, machine learning has emerged as a powerful method to overcome the roadblocks. Here, we propose a self-supervised machine learning approach to not only identify unknown phases but also unmix the overlapped chemical signals of individual phases with no need for user expertise in mineralogy. This approach leverages the guidance of gaussian mixture modelling (GMM) clustering fitted on an informative latent space of pixel-wise elemental data points modelled using a neural network autoencoder, and deconvolutes the overlapped chemical signals of phases using non-negative matrix factorisation (NMF). We evaluate the reliability and the accuracy of the new approach using two hyperspectral EDX datasets. The first dataset was measured from an intentionally fabricated sample, where seven known mineral particles are physically overlapping with each other as well as the substrate. Without any prior knowledge, the proposed approach successfully identified all major phases and recovered the original chemical spectra of the individual phases with high accuracy. In the second case, the dataset was collected from a potential vehicular source of particulate matter air pollution, where identification of the individual pollution particles is complicated by the complex nature of the sample. The approach once again was able to identify the potential Fe-bearing ultrafine particles and isolate the background-subtracted elemental signal. We demonstrate a robust approach that potentially brings a significant improvement of mineralogical and chemical analysis in a fully automated manner. In addition, the proposed analysis process has been built into a user-friendly Python code with graphical user interface (GUI) for ease of use by general users.