Nowadays, new deposits are rarely developed for ore extraction, which is why partially depleted mines continue to operate. Due to the decrease in the concentration of valuable material, e.g. copper ores in most mines, the mass of material to be processed increases in order to obtain the same amount of the end product. This also increases the amount of water and energy used to extract these raw materials.

One way to counteract this depletion is to pre-concentrate the material stream by sensor-based sorting. In this process, tailings, the processing of which contributes only a minimal share to the end product, are removed early in the process chain. Since the crushing and grinding of the rocks in the mining operation requires large amounts of energy, there is great potential for savings here. Pre-sorting the material stream thus enables a reduction in the resources expended per ton of end product, such as energy and water, as well as consumables and reagents in the later process steps.

The basis of such sorting is a sensor system that makes a prediction of the concentration of the desired mineral or element for each individual particle. Of the possible sensor technologies for sorting, ME-XRT (multi-energy X-ray transmission imaging) and LIBS (laser-induced breakdown spectroscopy) were selected because they can provide complementary compositional information. ME-XRT is a transmission technique and provides information on the total irradiated volume of the sample, but is limited in the accuracy of elemental information, as it only provides information correlated with the effective atomic number. LIBS, on the other hand, is limited in analysis to a comparatively small area of the surface, but can provide more accurate information about the composition of chemical elements. By fusing the data from both sensor systems, the gain in accuracy of the prediction of the concentration that can be achieved as a result should be demonstrated. The accuracy of the prediction of the concentration should be studied using conventional methods for each sensor technology separately as a reference value. Using machine learning (ML) methods, specifically deep neural networks, the data from the sensor technologies should be examined both individually and in combination (data fusion).

The accuracy of the prediction of the concentration of the desired mineral or element should then be used to derive the possible savings in water and energy with appropriate presorting. Furthermore, the usability of the sensor information in the overall geological picture should be investigated. This should be demonstrated using Chilean copper and iron ores as an example.

First, suitable samples were selected in the consortium. Two different copper ore samples from the Rafaela mine in Chile were selected. The first type contains predominantly copper oxides as a valuable mineral, while the second type contains predominantly copper sulfides as valuable minerals. The Swedish project partner geologically characterized these samples. In the further course, additional iron ore samples were selected from the La Esterella and Mariposa mines (both in Chile).

For both ME-XRT and LIBS, a calibration step was first necessary. For ME-XRT, this consists of measuring a variety of high-purity samples of different thicknesses. These high-purity calibration samples consist of chemical compounds such as polyoxymethylene (POM) or pure fused quartz, as well as pure elements such as carbon (graphite), aluminum, titanium, iron, copper or lead. For the calibration of the LIBS system, the rock samples were measured in their original form and as finely ground powder. With the help of the reference analysis carried out in the laboratory, the LIBS system could then be calibrated. For one part of the copper ore samples (training set for the creation of the calibration model), a broad laboratory analysis was carried out for almost all relevant elements. For a second part (test set for validation of the calibration model), only the copper content was determined. For the iron ore samples, only the iron content was determined analogously. The values of these analyses served as ground truth for the evaluation of the prediction accuracy of the copper concentration.

The LIBS spectra showed a strong matrix dependence, which is why independent calibration models had to be created for the predominantly copper oxide and the predominantly copper sulphide samples as well as for the iron ore samples. For ME-XRT and LIBS a validation of the respective calibration model was carried out with an independent test set of the respective ore group.

In addition, eleven mineral-diverse rock samples were selected as representative samples by the Swedish project partner, the Technical University in Luleå, to investigate the suitability of ME-XRT and LIBS for element mapping at the micrometer scale in order to quickly obtain basic geological information about the deposits. The measured data of the ores were analyzed using both the conventional methods and deep neural networks. Due to the data fusion of ME-XRT and LIBS with such neural networks, which has not yet been described in the literature, different data pre-processing steps and network architectures were first designed and tested.

Based on the prediction accuracy for the copper grade, the Chilean project partner, the University of Chile in Santiago, modeled the potential resource savings and applied them to the acquired data.

The samples from the Rafaela mine showed a very fine structure with grain sizes in the range of a few μm to a few 10 μm in the geological survey at the Technical University in Luleå. Therefore, neither ME-XRT (pixel size 800 μm) nor LIBS (size of laser spot approx. 100 μm) detects and analyses the individual underlying minerals. In combination with the averaging of many measuring points over the entire sample, very homogeneous results are achieved overall, which brings a great advantage for the subsequent sorting.

The investigations of the different 11 mineral rock samples have shown that LIBS can be used for fast element mapping and thus important basic information for new deposits can be gained. Furthermore, based on the specific spectral fingerprint of each deposit, LIBS could be used to classify the rock samples with respect to their deposit. ME-XRT quickly provides the proportion of elements in a certain atomic number range, in this case mainly iron and copper.

Due to the limited radiolucency with X-rays, the realistic grain sizes are in the range of a few centimeters. Depending on the composition, i.e. essentially the density and atomic number of the elements involved, the thickness that could still be easily radiated with 160 kV accelerating voltage of the X-ray tube was between approx. 30 mm for the iron ores of La Estrella and Mariposa with iron content > 60 % and correspondingly high density, and approx. 50 mm for the copper ores of Rafaela.

In the evaluation of the prediction of the copper content, the coefficient of determination R2 was used as a comparative variable. For the LIBS measurements, values between 85 % and 99 % were achieved for the copper ore samples, while R2 values between 36 % and 74 % were shown here for the ME-XRT measurements. LIBS achieves a higher accuracy in prediction, but is subject to statistical fluctuations with inhomogeneous samples and few measuring points on the surface, while ME-XRT does not achieve this local accuracy, but includes the entire volume of the sample and allows higher throughputs in sorting.

Three different deep neural network architectures were designed for concentration evaluation and prediction, combined with three different preprocessing steps. In each case, the data from the sensor technologies were examined individually, as well as the fusion of the data. Data fusion increased R2 from 12 % (ME-XRT) and 28 % (LIBS) to 46 % (ME-XRT+LIBS) in cross-validation, and from 21 % (ME-XRT) and 31 % (LIBS) to 34 % (ME-XRT+LIBS) when tested with the previously unseen test data set. Thus, the fusion of the two sensor modalities showed an increase compared to the individual data evaluated with the same methods. However, the accuracy of prediction using ML methods is lower than for conventional data evaluation. The reason for this is probably the small amount of training data available for machine learning methods in this project.

By using such a sorting technique in the transfer of the results to the operation of mines of different sizes, the modelling of the Chilean project partners was able to show a reduction in energy consumption per ton of end product of approx. 33 % for large and medium/small mine operations. In large mining operations, the potential for water savings is more than 23 %; this figure was not determined for medium/small mines because the water-intensive concentration steps are not implemented there.

Further studies must now show what improvement a significant increase in the size of the training data brings for the ML methods and whether a fusion of the conventionally evaluated data is possible.

Project partners

Fraunhofer EZRT is an internationally leading research and development center in the area of non-destructive monitoring along the entire materials value chain of the product life cycle, ranging from raw materials via production towards recycling. Fraunhofer EZRT is defining and advancing the state of the art in this area, especially by applying imaging X-Ray and magnetic resonance techniques as well as optical inspection technologies. The research areas include sensor systems, simulation for data acquisition, image processing for data enhancement and evaluation (metadata acquisition), system development, metrology as well as applications and training.


Luleå University of Technology (LTU) is Scandinavia’s northernmost technical university. Through the geographical proximity and close collaborative ties with the Nordic mining industry, LTU has succeeded in establishing a world-class expertise spanning the entire value chain of mining. Mining-related research at LTU has the aim to provide new and improved solutions for securing a sustainable raw material supply which is crucial for the development of the modern society. The ore geology research group at LTU contributes to the geological understanding of ore deposits by integrating multi-scale 3D geological and geophysical subsurface modelling with ore formation studies and micro-analytical resource characterization. This approach and the resulting increased scientific knowledge about ore deposits continues to contribute greatly to the improvement of exploration and mining efficiency.

Laser-induced breakdown spectroscopy (LIBS) is one of the fastest and advanced optical spectroscopic techniques for atomic characterization of material. Being a highly innovative German company; SECOPTA analytics GmbH unites the latest state-of-the-art photonic technologies with the newest development in LIBS spectroscopy. For more than 10 years, it is providing the most advanced spectroscopic measurement system in the field of industrial quality control and process analysis. With key product line namely FiberLIBS lab, FiberLIBS Inline, MopaLIBS  and Mineral LIBS; SECOPTA conducts PMI (positive material identification),fast recycling application( identification of low and high alloy Al,Steel, Ti scrap), investigation of surface coating and surface ablation, refractory bricks and liquid metal analysis, onsite construction material analysis as well as quantitative and qualitative analysis of bulk material flow containing minerals. Specifically for mining applications, the new and highly developed multielement analyzer ‘MineralLIBS’ system simultaneously characterizes different atomic composition of moving (conveyor belt speed up to 3m/s) raw minerals as well as monitors individual ingredient element (Cu, C, Si, Al, Fe, Ca, Na, K, S, Mg, Pb, H, Zn, Ti, N, P, Mn) content on site with acute analytical precision and up to the lowest concentration limit of 100 ppm.


Founded in 1842, the University of Chile is the oldest institution of education in Chile. Although the career of Mining Engineering has been imparted since 1853, the Department of Mining Engineering was formally created in 1964. According the QS world university rankings the University of Chile is among the top 10 institutions in the field of mineral engineering.  The Minerals and Metals Characterisation and Separation (M2CS) research group founded in 2012 by Dr. G. Montes-Atenas aims at producing solutions for the mining industry in the fields of minerals processing and extractive metallurgy.  Dr. Montes-Atenas research group is currently acting as the Chilean counterpart of the REWO-SORT project which is expected to lead the development of sorting technologies reducing largely the energy costs in concentrators worldwide. The focus of the project is devoted not only to increasing the separation efficiency of valuable material at early stages of the mineral processing, but also to improving the treatment capacity of sorting technologies.