2292 XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3
OUTLOOK
Despite their simplicity, ML techniques have been demon-
strated so far to be quite effective in predicting adsorption
energies. Yet, these models come with a few limitations.
The first one deals with their transferability. Let us assume
for instance to train a ML model on a quite reactive surface.
Will be this model restricted to the adsorption of molecules
in one specific adsorption site (i.e., on-top or hollow)? Of
course, one can retrain the model to include and study one
peculiar property. Yet, how much of the previous training
will the model forget ?A second crucial restriction concerns
the high computational cost over the training. As men-
tioned above, the latter is related to the query strategy, in
other words how to generate the configurations to train the
model. To remedy to these issues, we are currently testing
and combining transfer learning (TL) modeling (Ma et
al. 2020) with divide and conquer techniques in order to
extend the capability of the predictive model, by including
a broader array of realistic configurations, such as stepped
edges or grain boundaries at the surface, and include a wide
range of molecular adsorbates.
Figure 8 depicts our transfer learning (TL) strategy
over the ML training. To overcome the training bottle-
neck, we first train our system separately (i.e., starting from
the substrate, then the solution, eventually the interface,
Figure 8(a)). In a second step, the stored MLFFs for dif-
ferent materials may be interchangeably merged and then
used for further prediction purposes (Figure 8(b)). Finally,
our TLFF model is tested on similar systems to improve the
flexibility and then the transferability of the built potential.
The latter may be achieved for instance by monitoring on
the fly the radial distribution functions. Of note, TLFF
can be considered as the key step within the overall ALFF
procedure. Indeed, whenever the MLFF does not meet the
criterion concerning stability and robustness, new data
points are added to the initial data set to improve the TL
capability.
CONCLUDING REMARKS
Depending on the microscopic property of interest, long
time and large size scales are required to converge impor-
tant dynamical properties of mineral-water interfaces. Since
classical simulations fail in providing the interface reac-
tive picture whereas ab initio methods are computation-
ally demanding and limited to short times, MLFF models
represent nowadays the cutting-edge solution to gauge rare
reactive events at reasonable training cost. In this work, we
made use of MLFF techniques to unravel water layering
structures and dynamics upon minerals. Findings in this
direction may pave the way for more effective adsorption
models in flotation industry. For instance, a complete lack
of information concerning the flotation of ferrosilite still
exists, as well as on the interactions of the flotation reagents
with the mineral. In addition, being difficult to find the
mineral in the pure form to buy, most part of investigations
are only possible using theoretical methodologies as ML
modeling. Studies in this direction aimed at the discovery
of new material properties by the efficient investigation of
the configuration space as well as improving the transfer
Figure 7. MLFF simulation of the calcite/OA interface. Left: initial disordered configuration
concerning the collector conditioning. Right: final configuration with the oleate adsorbed via
carboxylate termination
OUTLOOK
Despite their simplicity, ML techniques have been demon-
strated so far to be quite effective in predicting adsorption
energies. Yet, these models come with a few limitations.
The first one deals with their transferability. Let us assume
for instance to train a ML model on a quite reactive surface.
Will be this model restricted to the adsorption of molecules
in one specific adsorption site (i.e., on-top or hollow)? Of
course, one can retrain the model to include and study one
peculiar property. Yet, how much of the previous training
will the model forget ?A second crucial restriction concerns
the high computational cost over the training. As men-
tioned above, the latter is related to the query strategy, in
other words how to generate the configurations to train the
model. To remedy to these issues, we are currently testing
and combining transfer learning (TL) modeling (Ma et
al. 2020) with divide and conquer techniques in order to
extend the capability of the predictive model, by including
a broader array of realistic configurations, such as stepped
edges or grain boundaries at the surface, and include a wide
range of molecular adsorbates.
Figure 8 depicts our transfer learning (TL) strategy
over the ML training. To overcome the training bottle-
neck, we first train our system separately (i.e., starting from
the substrate, then the solution, eventually the interface,
Figure 8(a)). In a second step, the stored MLFFs for dif-
ferent materials may be interchangeably merged and then
used for further prediction purposes (Figure 8(b)). Finally,
our TLFF model is tested on similar systems to improve the
flexibility and then the transferability of the built potential.
The latter may be achieved for instance by monitoring on
the fly the radial distribution functions. Of note, TLFF
can be considered as the key step within the overall ALFF
procedure. Indeed, whenever the MLFF does not meet the
criterion concerning stability and robustness, new data
points are added to the initial data set to improve the TL
capability.
CONCLUDING REMARKS
Depending on the microscopic property of interest, long
time and large size scales are required to converge impor-
tant dynamical properties of mineral-water interfaces. Since
classical simulations fail in providing the interface reac-
tive picture whereas ab initio methods are computation-
ally demanding and limited to short times, MLFF models
represent nowadays the cutting-edge solution to gauge rare
reactive events at reasonable training cost. In this work, we
made use of MLFF techniques to unravel water layering
structures and dynamics upon minerals. Findings in this
direction may pave the way for more effective adsorption
models in flotation industry. For instance, a complete lack
of information concerning the flotation of ferrosilite still
exists, as well as on the interactions of the flotation reagents
with the mineral. In addition, being difficult to find the
mineral in the pure form to buy, most part of investigations
are only possible using theoretical methodologies as ML
modeling. Studies in this direction aimed at the discovery
of new material properties by the efficient investigation of
the configuration space as well as improving the transfer
Figure 7. MLFF simulation of the calcite/OA interface. Left: initial disordered configuration
concerning the collector conditioning. Right: final configuration with the oleate adsorbed via
carboxylate termination