Application of Grade Control Model updating by Compositional Sequential Ensemble Filtering

A key requirement for a streamlined mine-to-mill process is the characterization of the spatial distribution of geometallurgical properties within the mineral deposit, such as ore grade or proportions of deleterious components or the gangue mineralogy.

Due to the limited amount of information available during exploration and grade-control, underlying spatial models, such as the resource or grade-control model, are associated with uncertainty and do not represent reality and production forecast of ROM ore.

One opportunity to decrease these errors is to assimilate grade-control monitoring data, which are obtained by online sensor technology in many operations, into resource or grade-control models. This concept leads to a closed-loop reconciliation system, that generates updated models on real-time, providing up to date information for decision makers in mine planning and operations control. Recently, univariated approaches of data assimilation have been documented and successfully operationally implemented.

The extension to the multivariate case comes with additional challenges. Certain geometallurgical properties, like chemical or mineral compositions, are measured online based on the ratio between percentages and do not represent absolute values. Classical data assimilation outcomes need to be corrected to account for mass conservation of each component, while keeping their physical relations, such as the stability of mineral assemblages, the total sum constraint or positivity. Therefore, new methods for geostatistical modeling and information reconciliation are needed that can account for such issues in a natural way.

This contribution presents a new compositional based data assimilation approach. This supersedes the problems of positivity preservation and mass conservation by working with log-ratios of components. For optimal performance, a flow anamorphosis transformation is used to introduce normality to observations and model variables. This step is necessary in most data assimilation methods.

The contribution presents the methodology and demonstrates it applied in a 3D case study from a Bauxite deposit.