Learning a Meaningful Representation with Prior Domain Knowledge Applied to X-ray Sensing in Geology

Matthew Dirks

Abstract: Mining companies typically process all the material extracted from a mine site using processes which are extremely consumptive of energy and chemicals, but by sorting valuable ore from waste rock we can reduce costs and help preserve the environment. X-ray Fluorescence sensors produce spectra data that can be used to quantify element concentrations in rock samples. Our goal is to train a model to convert X-ray Fluorescence spectra into element concentrations automatically using a small dataset of rock samples that have ground truth associated with them via geochemical assay. In machine learning, when not enough data is available (often due to increasingly complex models) or the quality of the data is insufficient, then prior domain knowledge from experts can be incorporated to guide the learner. Many applications have access to prior domain knowledge, but it's not always clear how to utilize it. We introduce a system that encodes X-ray Fluorescence spectra into a meaningful low-dimensional representation via a variation of an AutoEncoder where the decoder is a forward simulator that incorporates prior domain knowledge. We train the AutoEncoder unsupervised, then swap the decoder with a module that outputs the desired prediction targets via supervised learning. We evaluate our method with and without pre-training and compare to the baseline quantitative method. We show that our method better estimated element concentrations using the pre-trained AutoEncoder with incorporated prior domain knowledge.

TL;DR: We predict element concentrations in rocks via a representation of X-ray Fluorescence sensor data that is learnt using a variation of an AutoEncoder where the decoder is a forward simulator that incorporates prior domain knowledge.

Keywords: AutoEncoder, Prior Knowledge, Domain Knowledge, Representation, XRF, Semi-supervised Learning