1156
Optimizing Physical Separation Through Machine Learning
Based Mineralogy Classification
Marion Nicco, David Bednarski, Maud Herbelin, Thomas Wallmach
Eramet Ideas
ABSTRACT: Monitoring the performance of mineral processing equipment is crucial for cost effective
production. Using the example of a heavy minerals sand mine, a software based on open-source components was
developed to determine the mineralogy of production samples from photographs. The program processes images
in two steps: instance segmentation of the particles using Segment Anything Model (SAM) then classification
using machine learning clustering and image recognition techniques. At the interface between research and
industry, this web application is designed for daily use by metallurgists to optimize physical separation.
Keywords: Segment Anything Model (SAM), Mineral processing, Mineralogy
INTRODUCTION
The mining industry, a dynamic sector fueling global
economies, relies on an equilibrium of geological knowl-
edge, technological innovation, and operational efficiency.
At its base lies geology and in particular mineralogy—the
scientific study of minerals and their properties—a disci-
pline that stands as the foundation for mining operations
worldwide.
Traditionally the monitoring of mineral physical sepa-
ration is largely done using traditional X-ray Fluorescence
(XRF) from stream samples and when looked at, mineral-
ogy is measured through SEM based automated mineral-
ogy methods, but these require sample preparation. Online
analyzers can be tied with the separation equipment. These
online analyzers are either XRF (Hołyńska et al., 1985)
or more recently Laser Induced Breakdown Spectroscopy
(LIBS) (Khajehzadeh et al., 2016) which are extremely
expensive solutions and still only gives elemental informa-
tion and an estimation of the mineralogy, which often is
not sufficient to efficiently control the separation process
(Berry et al., 2008 Tomanec et al., 2014 Koh et al., 2024).
Optical microscopy analysis remains the reliable
method to instantly examine and assess mineralogy (De
Castro, 2022). Done by hand this process is both time-con-
suming and sensitive to subjectivity of the operator (Gomes
et al., 2013 Iglesias et al., 2019). With the emergence of
advanced technologies, there has been a growing interest
in automating the segmentation and classification from
optical microscopy images (transmitted or reflected) to
enhance efficiency and accuracy of mineral process (Nellros
and Thurley, 2022 Koh et al., 2024) and overall get better
understand of ore to anticipate downstream performance
(Köse et al., 2012 Gomes et al., 2013 Iglesias et al., 2019
Santoro et al., 2022).
Image segmentation can be done using standard algo-
rithms and artificial intelligence, available as open source
(notably scikit-image from scikit-learn, 2011) and com-
mercial packages (ImageJ, MATLAB, Fiji etc.). This paper
presents the steps taken for building using newly released
Segment Anything (SAM) algorithm from Meta (Kirillov
et al., 2023) as a tool for an application for segmenting and
identifying minerals, as well as examples of its performance.
Optimizing Physical Separation Through Machine Learning
Based Mineralogy Classification
Marion Nicco, David Bednarski, Maud Herbelin, Thomas Wallmach
Eramet Ideas
ABSTRACT: Monitoring the performance of mineral processing equipment is crucial for cost effective
production. Using the example of a heavy minerals sand mine, a software based on open-source components was
developed to determine the mineralogy of production samples from photographs. The program processes images
in two steps: instance segmentation of the particles using Segment Anything Model (SAM) then classification
using machine learning clustering and image recognition techniques. At the interface between research and
industry, this web application is designed for daily use by metallurgists to optimize physical separation.
Keywords: Segment Anything Model (SAM), Mineral processing, Mineralogy
INTRODUCTION
The mining industry, a dynamic sector fueling global
economies, relies on an equilibrium of geological knowl-
edge, technological innovation, and operational efficiency.
At its base lies geology and in particular mineralogy—the
scientific study of minerals and their properties—a disci-
pline that stands as the foundation for mining operations
worldwide.
Traditionally the monitoring of mineral physical sepa-
ration is largely done using traditional X-ray Fluorescence
(XRF) from stream samples and when looked at, mineral-
ogy is measured through SEM based automated mineral-
ogy methods, but these require sample preparation. Online
analyzers can be tied with the separation equipment. These
online analyzers are either XRF (Hołyńska et al., 1985)
or more recently Laser Induced Breakdown Spectroscopy
(LIBS) (Khajehzadeh et al., 2016) which are extremely
expensive solutions and still only gives elemental informa-
tion and an estimation of the mineralogy, which often is
not sufficient to efficiently control the separation process
(Berry et al., 2008 Tomanec et al., 2014 Koh et al., 2024).
Optical microscopy analysis remains the reliable
method to instantly examine and assess mineralogy (De
Castro, 2022). Done by hand this process is both time-con-
suming and sensitive to subjectivity of the operator (Gomes
et al., 2013 Iglesias et al., 2019). With the emergence of
advanced technologies, there has been a growing interest
in automating the segmentation and classification from
optical microscopy images (transmitted or reflected) to
enhance efficiency and accuracy of mineral process (Nellros
and Thurley, 2022 Koh et al., 2024) and overall get better
understand of ore to anticipate downstream performance
(Köse et al., 2012 Gomes et al., 2013 Iglesias et al., 2019
Santoro et al., 2022).
Image segmentation can be done using standard algo-
rithms and artificial intelligence, available as open source
(notably scikit-image from scikit-learn, 2011) and com-
mercial packages (ImageJ, MATLAB, Fiji etc.). This paper
presents the steps taken for building using newly released
Segment Anything (SAM) algorithm from Meta (Kirillov
et al., 2023) as a tool for an application for segmenting and
identifying minerals, as well as examples of its performance.