7
and its python interface as well as automated code genera-
tion capabilities that offers a high-performance computing
environment for solving PDEs through FEM [24].
THE HYBRID MODEL
The hybrid model accelerates the coupled model for visu-
alization and is used for generating different scenarios in
a VR environment. Hybrid models are used to accelerate
computationally expensive numerical solutions of initial
boundary value problems. In [25], in the context of mul-
tiscale FE2 simulations, a neural network based surrogate
at the micro-scale is used to achieve a speed-up of a factor
of more than 5000 compared to a reference FE2 compu-
tation. The applied scientific machine learning, known as
operator learning [26], approximates the mapping between
infinite dimensional spaces that arise from physical mod-
els expressed as PDEs. This operator is a generalization
of the neural network as the input and outputs are func-
tions and not vectors. Some state-of-the art architectures
as reported in the literature include deep operator network
(DeepONet) [27] and Fourier neural operator [28].
A DeepONet has been chosen for the MOVIE project.
The encoder-decoder network structure as in DeepONet
encodes the input function using a finite dimensional vec-
tor. The encoded input is mapped to a finite dimensional
output. This is followed by decoding the finite dimensional
output to an output function in the infinite dimensional
function space [29]. This approach has the structure of
many numerical schemes such as FEM, which is used in
the geomechanical model. In FEM, the encoding is the
Galerkin projection, and the mapping is done by a numeri-
cal scheme and the decoding is done by finite element basis.
The manually designed algorithm and chosen numerical
discretization are replaced by a data-driven approach, allow-
ing for encoding and reconstruction directly from data.
The steps involved in creating a hybrid model are data
generation, model training and evaluation. In the data gen-
eration phase, data in the form of examples are generated
from simulating different scenarios. Then a neural opera-
tor is trained to make predictions. The predictions are then
compared to the reference simulations and evaluated. This
means that the hybrid model considers the examples gener-
ated from coupled simulations and tries to approximate the
coupled simulation using a neural network.
This method and its extensions based on reduced basis
like proper orthogonal decomposition DeepONet [30],
helps in reducing the computational effort and enables real
time updates of the model. This is essential for a digital
twin, as it allows simulation results from real-time data
collected at the mine to be promptly updated and visual-
ized in the virtual laboratory, minimizing any significant
time lag.
THE DATA TRANSFER OF REAL TIME
DATA TO THE DIGITAL TWIN OF THE
“FLB REICHE ZECHE”
In recent years, the concept of digital twins has become
increasingly important in the mining industry. One of the
reasons for this is that these systems can help to optimize
various processes and solve complex problems [31], [32].
Nevertheless, digital twins are currently used to a limited
extent within the industry and in most cases only relate to
individual elements of the process chain, such as conveyor
systems and processing plants [33], [34]. They are used as
decision-making tool, for the management and coordina-
tion of processes or for the monitoring and maintenance
of machines and facilities [31]. In contrast, the digital twin
of the underground laboratory in the “FLB Reiche Zeche,”
will act as a visualization of the hybrid model and be used
for a more holistic representation of the selected under-
ground laboratory.
An initial illustration of the later representation of
the digital twin into the target software LiquidEarth One
developed by Terranigma Solutions GmbH in Germany is
shown in Figure 7. The objective is to provide the option to
visualize the four individual models and the fully coupled
model (computing time optimization of the hybrid model.)
using the VR capability of LiquidEarth One.
To continuously update the digital twin with real time
data from the underground laboratory, the data collected
by the different sensors install is compiled and transferred
using the DECOMDA system. DECOMDA is a decen-
tralized communication and data acquisition system for
underground operations that uses stationary and mobile
data loggers to collect and temporarily store data. This data
is then transferred to data collectors that are either vehicle-
or person-bound and help to transport the data to central
areas of the mine, where it is uploaded to higher systems to
finally evaluate and visualize it [35]. DECOMDA creates
the interface between digital tools and geoscientific applica-
tions as a part of the physical-to-virtual connection [36].
The measurement data obtained from the underground
laboratory is used to validate and improve the various indi-
vidual models. The resulting information is then incorpo-
rated through the hybrid model, providing an up-to-date
representation of the underground conditions as a part of
the digital twin.
and its python interface as well as automated code genera-
tion capabilities that offers a high-performance computing
environment for solving PDEs through FEM [24].
THE HYBRID MODEL
The hybrid model accelerates the coupled model for visu-
alization and is used for generating different scenarios in
a VR environment. Hybrid models are used to accelerate
computationally expensive numerical solutions of initial
boundary value problems. In [25], in the context of mul-
tiscale FE2 simulations, a neural network based surrogate
at the micro-scale is used to achieve a speed-up of a factor
of more than 5000 compared to a reference FE2 compu-
tation. The applied scientific machine learning, known as
operator learning [26], approximates the mapping between
infinite dimensional spaces that arise from physical mod-
els expressed as PDEs. This operator is a generalization
of the neural network as the input and outputs are func-
tions and not vectors. Some state-of-the art architectures
as reported in the literature include deep operator network
(DeepONet) [27] and Fourier neural operator [28].
A DeepONet has been chosen for the MOVIE project.
The encoder-decoder network structure as in DeepONet
encodes the input function using a finite dimensional vec-
tor. The encoded input is mapped to a finite dimensional
output. This is followed by decoding the finite dimensional
output to an output function in the infinite dimensional
function space [29]. This approach has the structure of
many numerical schemes such as FEM, which is used in
the geomechanical model. In FEM, the encoding is the
Galerkin projection, and the mapping is done by a numeri-
cal scheme and the decoding is done by finite element basis.
The manually designed algorithm and chosen numerical
discretization are replaced by a data-driven approach, allow-
ing for encoding and reconstruction directly from data.
The steps involved in creating a hybrid model are data
generation, model training and evaluation. In the data gen-
eration phase, data in the form of examples are generated
from simulating different scenarios. Then a neural opera-
tor is trained to make predictions. The predictions are then
compared to the reference simulations and evaluated. This
means that the hybrid model considers the examples gener-
ated from coupled simulations and tries to approximate the
coupled simulation using a neural network.
This method and its extensions based on reduced basis
like proper orthogonal decomposition DeepONet [30],
helps in reducing the computational effort and enables real
time updates of the model. This is essential for a digital
twin, as it allows simulation results from real-time data
collected at the mine to be promptly updated and visual-
ized in the virtual laboratory, minimizing any significant
time lag.
THE DATA TRANSFER OF REAL TIME
DATA TO THE DIGITAL TWIN OF THE
“FLB REICHE ZECHE”
In recent years, the concept of digital twins has become
increasingly important in the mining industry. One of the
reasons for this is that these systems can help to optimize
various processes and solve complex problems [31], [32].
Nevertheless, digital twins are currently used to a limited
extent within the industry and in most cases only relate to
individual elements of the process chain, such as conveyor
systems and processing plants [33], [34]. They are used as
decision-making tool, for the management and coordina-
tion of processes or for the monitoring and maintenance
of machines and facilities [31]. In contrast, the digital twin
of the underground laboratory in the “FLB Reiche Zeche,”
will act as a visualization of the hybrid model and be used
for a more holistic representation of the selected under-
ground laboratory.
An initial illustration of the later representation of
the digital twin into the target software LiquidEarth One
developed by Terranigma Solutions GmbH in Germany is
shown in Figure 7. The objective is to provide the option to
visualize the four individual models and the fully coupled
model (computing time optimization of the hybrid model.)
using the VR capability of LiquidEarth One.
To continuously update the digital twin with real time
data from the underground laboratory, the data collected
by the different sensors install is compiled and transferred
using the DECOMDA system. DECOMDA is a decen-
tralized communication and data acquisition system for
underground operations that uses stationary and mobile
data loggers to collect and temporarily store data. This data
is then transferred to data collectors that are either vehicle-
or person-bound and help to transport the data to central
areas of the mine, where it is uploaded to higher systems to
finally evaluate and visualize it [35]. DECOMDA creates
the interface between digital tools and geoscientific applica-
tions as a part of the physical-to-virtual connection [36].
The measurement data obtained from the underground
laboratory is used to validate and improve the various indi-
vidual models. The resulting information is then incorpo-
rated through the hybrid model, providing an up-to-date
representation of the underground conditions as a part of
the digital twin.