Geofingerprinting of Coltan Using Handheld Spectroscopic Devices

Following the enactment of the Dodd-Frank Act in 2010, specifically Sect. 1502, US companies have been required to report utilizing conflict minerals from the Democratic Republic of Congo (DRC). The conflict mineral coltan, an ore consisting of elements tantalum and niobium, is central to this issue and engenders the need to track and trace the mineral's supply chain. X-ray fluorescence (XRF) and laser-induced breakdown spectroscopy (LIBS) have been used, in combination with both unsupervised and supervised machine learning, to accurately classify coltan samples with known provenances. Sample spectra were first used as input data into unsupervised machine learning clustering algorithms, upon which dendrogram and constellation plots were generated. The classification achieved via unsupervised machine learning provided the proof of concept necessary to further investigate classification using supervised machine learning algorithms. The sample's raw spectra were then used to train a supervised machine learning algorithm, consisting of a voting classifier relying on the results from random forest classifier (RFC), linear regression classifier (LRC), support vector classifier (SVC), and multi-layer perceptron classifier (MLPC). The classification achieved using raw spectra was able to achieve accuracies up to similar to 97%. The samples' raw spectra were pre-processed using principal component analysis (PCA), and the pre-processed data was fed into the same supervised machine learning classifier described above. Accuracies of similar to 98% and similar to 96%, respectively, were achieved. When reviewing the predicted classifications arising from the use of these two different types of spectra, specifically reviewing the confidence score associated with each predicted provenance classification, it was possible to account for the incorrect provenance classifications returned by the voting classifier. If the predicted provenance and associated confidence score obtained via each spectra type was compared to the resulting provenance prediction and confidence score obtained by the other spectra type, and only the prediction with the higher associated confidence score was used, classification accuracies of 100% could be achieved.