Integrated Approach to Identify Variables for the Prediction of Metallic Content in a Slag Heap using Time-Domain ERT and SIP Laboratory Measurements

There are two main drivers to integrate former metallurgical residues or mine waste into the circular economy through an efficient resource recovery. First, the economic driver, which considers land and resource recovery of high value elements such as critical raw materials. Secondly, the environmental and human health driver, as these types of residues might be a potential source of pollution. In both scenarios, there is a need to improve the characterization of past metallurgical sites and to locate and quantify materials of interest. To this aim, mostly geoelectric methods applied in the laboratory and/or in the field have been used and they are often complemented with geochemical or mineralogical studies. In this contribution, we present an approach that integrates a 3D Electrical Resistivity Tomography (ERT) and Induced Polarization (IP) acquisition in the field, lab measurements of ERT, IP and Spectral Induced Polarization (SIP) in several samples together with chemical analysis, to predict the metallic content in a slag heap from a former iron and steel factory located in Belgium. The samples were collected at locations targeting the observed geophysical anomalies. We first look for correlations between geophysical lab measurements and the chemical analysis to identify the variables which could potentially have a larger impact or control in the metallic content. Second, we use the geophysical lab measurements to improve the deterministic constrained inversion carried out for the 3D field data. Third, we use a supervised learning algorithm - Gaussian Process Classification (GPC)- to predict the metallic content of the slag heap from the 3D inverted resistivity/chargeability model. Overall, we found that variations in the chargeability are correlated with changes in iron, calcium and silicon content. Additionally, the GPC represents a suitable algorithm to integrate the uncertainty in the prediction results as well as the uncertainty that arises from the direct comparison of field and lab data. Finally, this methodology which integrates geophysical field data, targeted sampling, lab measurements and supervised learning can be applied in broader contexts where such data are available.