Optim Eng

DOI 10.1007/s11081-016-9321-6

Stope optimization with vertical convexity constraints

Gonzalo Nelis1 • Michel Gamache2 • Denis Marcotte3 •

Xiaoyu Bai3

Received: 8 March 2015 / Revised: 1 January 2016 / Accepted: 13 March 2016

Ó Springer Science+Business Media New York 2016

Abstract A new algorithm for the optimal stope design problem is proposed. It is
based on a previous methodology developed by Bai et al. (Comput Geosci
52:361–371, 2013a) where a cylindrical coordinate system is used to define
geomechanical restrictions and to find the optimal stope around an initial raise. The
new algorithm extends this work using an integer programming formulation and a
new set of constraints, aimed to solve geomechanical issues present on the original
methodology. The new formulation is tested on two synthetic and one real deposits.
An economic, geomechanical and feasibility analysis is performed, comparing the
new results with Bai’s methodology. This methodology achieves better stope
designs in terms of geomechanical stability and wall regularity, generating feasible
stopes for real use. It also allows further extensions to incorporate other geometrical
constraints in order to obtain more regular stope designs.

Keywords OR in mining  Stope optimization  Underground mining  Sublevel

stoping

& Gonzalo Nelis

[email protected]
1
Department of Mining Engineering, Delphos Mine Planning Laboratory and Advanced Mining
Technology Center, University of Chile, Av. Tupper 2069, 8370451 Santiago, Chile
2
GERAD Research Center and Department of Mathematics and Industrial Engineering, École
Polytechnique, P.O. Box 6079, Station Centre-Ville, Montreal, QC H3C 3A7, Canada
3
Department of Civil, Geological and Mining Engineering, École Polytechnique, P.O. Box 6079,
Station Centre-Ville, Montreal, QC H3C 3A7, Canada

123
 G. Nelis et al.

1 Introduction

The main objective of underground mining is to achieve the highest profit from the
orebody while maintaining a safe operation. To achieve this, both economical and
geomechanical factors have to be considered for the selection of the mining method
and the design of the associated stopes. One of the most widespread underground
mining systems is ‘‘Sublevel Stoping’’, in which an open stope is created with
dimensions according to geomechanical conditions of the rock mass, the selectivity
desired and profitability of the stope. Figure 1 shows a scheme of this method.
Several methods have been proposed to address the underground design problem
and they can be classified in 1D, 2D and 3D methods. Among the 1D and 2D
methods, the most relevant are the dynamic programming algorithm (Riddle 1977)
and the branch and bound technique (Ovanic and Young 1995), which find the
optimal design but they do not consider the three dimensional nature of the problem.
On the other hand, the most relevant 3D methods are the floating stope algorithm
(Alford 1996), the octree division algorithm (Cheimanoff et al. 1989), the maximum
value neighborhood method (Ataee-Pour 2000), and simulated annealing techniques
(Manchuk and Deutsch 2008). All these methods address the three dimensions, but
they cannot incorporate real geomechanical restrictions in their formulation. Some

Fig. 1 A scheme of the sublevel stoping method. The different figures show the progression from the
initial raise and the ore blasting to the final open stope

123
Stope optimization with vertical convexity constraints

of these methods may incorporate an additional step to test the stability of the
current solution in order to generate stable designs, but they cannot address the
geomechanical restrictions directly. A comprehensive review about these methods
can be found in Ataee-Pour (2005).
Other methodologies are based on evaluating different locations of predefined
stopes to find the best global layout. Topal and Sens (2010) propose an algorithm to
find the optimum stope layout based on different possible stope sizes and using the
economic value of these stopes to choose the best layout. However this
methodology only generates a single mine layout when the optimum must be
chosen from all the possible layouts for the given stope sizes. Sandanayake et al.
(2015) address this issue with a heuristic approach which considers a set of possible
stope layouts to find the optimum. However, this approach still depends on the
initial stope sizes used in the algorithm and does not consider the rock pillars
between stopes.
Recent progress have been made on this subject for the case of sublevel stoping.
In Bai et al. (2013a), the authors find the stope design that maximizes the profit
obtained. Their approach is based on a transformation from the Cartesian coordinate
system to a cylindrical system centered around the initial raise. The authors could
define the optimal stope incorporating some geomechanical constraints as the
footwall and hangingwall angles, the minimum stope width that ensures the flow of
the blasted ore towards the raise, and the maximum distance from the raise to
control the size of the final stope. This approach was extended to include multiple
raises (Bai et al. 2014) and to incorporate drift development in the optimization
process (Bai et al. 2013b). This problem can be solved using a network structure
and an efficient maximum flow algorithm (Picard 1976).
In deposits with regular geometries, with a homogeneous grade distribution, this
methodology generates feasible stope designs. However, on irregular deposits, with
scattered mineralization, the stope designs may present stability problems with
irregular contours, despite the imposed restrictions of footwall and hangingwall
angles.
The stope stability is defined based on the rock mass properties such as uniaxial
or triaxial compressive strength, rock mass condition, presence of faults or
structures, etc. (Bieniawski 1989; Laubscher 1990). To ensure the stability of a
given stope it is necessary to define its shape, maximum dimensions, footwall and
hangingwall angles and the global stope inclination (Mathews 1980). The
methodology proposed by Bai successfully incorporates the footwall and hanging-
wall angles and the stope maximum dimensions, but in some cases the stopes
present shapes that reduce their global stability. Figure 2 illustrates such a stope
obtained with Bai’s approach where stability problems may occur, resulting in
increased ore dilution. This dilution comes from the walls of the stope in its
midsection, where a failure may occur given its non-convex shape, depending on the
quality of the rock mass. If the wall rock is blocky, with multiple sets of joints and
with poor surface quality (GSI lower than 40 according to Hoek et al. 2002), the
cohesion between this section and the wall rock will be low, so the failure
probability would be high. The quality of the blasting will also affect the probability
of failure of this section: a poor blasting technique will damage the surrounding

123
 G. Nelis et al.

Fig. 2 Stope generated by Bai’s

methodology, with stability
issues. The stope is 12 m wide
and 25 m tall approximately. YZ
vertical section at X = 30

walls of the stope, decreasing the strength of the wall rock and therefore, increasing
the probability of failure. Furthermore, a high horizontal stress condition will
generate a tension zone in this section with a negative impact on its stability. Given
this scenario, a methodology which fulfills the maximum stope dimensions and also
avoids this kind of geometries would be useful for some rock mass and mining
conditions.
In the next section, we briefly review the method used to measure stope stability.
We show how the lack of convexity along the stope sides influence negatively the
stability. We recall the main points of the approach used by Bai et al. (2013a). Then
we describe the new linear program used to enforce a particular form of convexity
of the stopes in the vertical direction. Finally, the results obtained with and without
the convexity constraints are compared on a few simple synthetic deposits and a real
deposit.

2 Methods

2.1 Stope stability determination

To study the stope stability, we use the strength factor defined as:
Rock mass strength
SF ¼ : ð1Þ
Induced stress
Figure 3 presents the SF obtained from a 2D stability analysis of the stope central
cross-section.
The analysis is done using Examine2D (Roclab, 2014). The software implements
2D boundary elements analysis (Crouch and Starfield 1983) to calculate the stress
induced in the rock mass around the excavation. It assumes a linear, elastic,
continuous, isotropic and homogeneous rock mass and, of course, it neglects the
third dimension. Despite these strong assumptions and simplifications, the stability

123
Stope optimization with vertical convexity constraints

Fig. 3 Stability analysis of

stope generated by Bai’s
methodology, with stability
issues. The midsection presents
possible failure by tension

Fig. 4 Stability analysis of a stope with the same dimensions as Fig. 3, but with a straight midsection.
The previous tension zone presents a better stability

analysis remains useful to compare relative stability of different stope designs in a

given rock mass.
For simplicity, the failure criterion to calculate SF was Mohr–Coulomb
(Coulomb 1776), but any other failure criterion could be used. The midsection
presents a possible failure by tension zone. This section will flow inside the stope,
diluting the ore and reducing the stope profit. Hence, this design could be not
optimal when considering the additional dilution. For comparison, a stope with the
same dimensions is presented in Fig. 4 but with a straight midsection.

2.2 Model formulation

The proposed model is based on the ultimate pit problem or final pit problem in
open pit mining (Lerchs and Grossmann 1965). The objective is finding the part of
the orebody which maximizes the profit obtained when it is extracted and processed.
To achieve this objective, the original orebody is discretized in small blocks with
different attributes such as metal content, extraction cost, processing cost, tonnage,

123
 G. Nelis et al.

coordinates, etc. From these attributes, an economic block model is defined using a
profit function (Lane 1988), where each block has a value representing the net profit
of extracting that particular block. It is important to mention that the set of blocks
extracted must ensure the pit wall stability, typically defined as a wall pit angle. For
this reason, geometrical precedence constraints are defined between blocks to
maintain the required angle. Formally, this problem can be defined as a linear
integer program (IP; Lerchs and Grossmann 1965):
X
max x i pi ; ð2Þ
i2N

s:t: xi  xj 8ði; jÞ 2 P i ; ð3Þ

xi 2 f0; 1g 8i 2 N: ð4Þ
The variable xi equals 1 if the block i in N is extracted and 0 otherwise. The
parameter pi is the extracting value of block i. The restriction set is derived from the
precedence constraints between blocks to maintain the wall pit angle. The set P i
contains all ordered pairs ði; jÞ where block j must be extracted before block i. Fi-
nally, the objective is to find the set of blocks that maximizes the profit, subject to
every block in this set must be extracted towards the surface fulfilling the prece-
dences between blocks.
Similarly, in the sublevel stoping mining system, a vertical raise is drilled in
order to create enough space to blast the ore inside the stope. The blasted material
fills the raise and falls down by gravity towards the drawpoints located in the bottom
of the stope. Sequential blasts are performed to extract all the material in the stope
taking advantage of the new space created from former blasts. Therefore, the role
played by this initial raise is analogous to the surface in open pit mining: all the
blocks must move towards the raise/surface in order to be extracted. Hence, the
precedences between blocks should link every block to the initial raise. This is the
base for the cylindrical coordinate system transformation proposed by Bai et al.
(2013a).

2.3 Cylindrical block model

In Bai et al. (2013a), an algorithm to transform a conventional block model (in the
Cartesian coordinate system) into a cylindrical block model is proposed in order to
emulate the role played by the ground surface in open pit mining with the initial
raise in sublevel stoping. The raise is defined as the center of the cylindrical system,
and a set of regular blocks are defined surrounding this raise from a discretization of
Dr; Dh and Dz defined by the user. This transformation presents a big advantage: to
extract any block of the block model, diagonal links can be defined in the cylindrical
system towards the raise, and this structure of links can be represented in the three-
dimensional space to generate the geometry shown in Fig. 5.
This figure shows that any block to be extracted forces all its predecessors
towards the raise to be extracted too, emulating the desired behavior. Furthermore,

123
Stope optimization with vertical convexity constraints

Fig. 5 Illustration of the cylindrical coordinate system; a horizontal precedences in the (r; h) plane, b the
same blocks as in a shown in their real spatial position, and c precedences in the vertical plane to generate
the footwall and hangingwall constraints. Shaded blocks are the blocks to remove to extract block A.
Black point in the center represents the raise

using this transformation, the footwall and hangingwall angles can be defined
directly with diagonal links as shown in Fig. 5c. Also, the discretization of Dr; Dh
and Dz can be chosen based on the footwall and hangingwall angles and the stope
width parameter (ymax
R ) to allow a steady flow of material towards the raise. This
problem can be solved using a network structure and efficient maximum flow or
other algorithms. After solving the problem in the cylindrical system, a back-
transformation is necessary to get the optimum stope design in terms of the original
(Cartesian) block model.
In this paper we will use the same transformation proposed by Bai et al. (2013a),
with the same precedences of footwall and hangingwall angles and the minimum
stope width. The raise location is considered known, but it can be optimized as in
Bai et al. (2014). This is the base used to improve the algorithm with the new
constraints presented in the next section.

2.4 Stope convexity constraints

It is not possible to avoid the irregular shapes in the stope outer section by imposing
vertical upward and downward precedences between blocks in the cylindrical
coordinate system as this would impose to take the entire column of blocks as soon
as a block is part of the stope. For this reason, the restriction must be imposed in a
different way. In Fig. 6a, a single column from a block model is presented. To avoid
non-convex geometries in the vertical direction, we must consider that if a block
from this column is extracted, and the block immediately below is not, then all the
blocks below must not be extracted. Equivalently, if a block is extracted and the
block immediately above is not, then all the blocks above must not be extracted. For

123
 G. Nelis et al.

Fig. 6 A single column of a

zmax A
block model
B i

C j

D
CA
E
k ∈ Cj
... CD ...

zmin n n

(a) (b)

instance, if block A is extracted and block B is not, all the blocks below B must not be
extracted in order to avoid the generation of non-convex geometries. The advantage
of putting the restrictions in this way is that it does not interfere with either the
footwall or the hangingwall constraints because it does not force the vertical
extraction in all situations. Note that in the horizontal plane, the stope can be non-
convex. However, this is not expected to cause any potential stability problems.
Although it is possible to enforce full convexity, this is not attempted here as it would
constrain the solutions to unrealistically smooth stopes perfectly symmetric around
the raise. Hereafter, we use the term convexity as a shortcut for vertical convexity.
For the implementation of this restriction, let define N the set of blocks. Let
T
N ¼ fi 2 N jzi ¼ zmax g be the set of all blocks located on the top level of the block
B
model and similarly let N ¼ fi 2 N jzi ¼ zmin g be the set of all blocks located on
the bottom level of the block model. Let define a column as the set of all blocks with
the same r and h coordinates, or the same x and y coordinates in the Cartesian
system. Let Ci ¼ fj 2 N j ri ¼ rj ; hi ¼ hj ; and zi [ zj g be the set of all blocks
below the block i in a single column. Finally, let C1i ¼ fj 2 Ci j zj ¼ zi  Dzg be the
set that includes only the block located immediately below i in Ci (e.g., in Fig. 6a,
C1A ¼ fBg). With these definitions, the IP to force convexity is shown:
X
max pi x i
i2N ð5Þ
s:t: xi  xj 8ði; jÞ 2 P i ;
B
x j þ aj  x i 8i 2 N nN ; j 2 C1i ; ð6Þ

T
x k  1  aj 8j 2 N nN ; 8k 2 Cj ; ð7Þ

xi ; ai 2 f0; 1g 8i 2 N : ð8Þ

The decision variable xi equals 1 if the block i 2 N is extracted and 0 otherwise,

and the coefficient pi is its extracting value. In constraint (6), j represents the block

123
Stope optimization with vertical convexity constraints

immediately below i in the block model (see Fig. 6b). The new binary variable aj
equals 1 if all the blocks k 2 Cj must not be extracted and 0 otherwise (i.e., they can
be extracted).
To clarify this concept, an example using Fig. 6a is shown where block C is
extracted (i.e., xC ¼ 1) and D is not (xD ¼ 0). Applying these restrictions to the
block C in this column :
xD þ aD  xC ; ð9Þ

x k  1  aD 8k 2 CD ¼ fE; . . .; ng: ð10Þ

In this case, if block C is extracted, and at the same time, D is not extracted, restriction
(9) forces the variable aD to be 1. If this variable is 1, the right side in restriction (10) is
0, forcing all the blocks k 2 CD to be 0, or, in other words, all the blocks below D must
not be extracted (i.e., xE ¼ 0; . . .; xn ¼ 0). Both restrictions combined avoid the non-
convexity issues presented before, ensuring a better stope shape.
Let’s now consider the blocks above a certain block. For example, let’s consider
that block D is extracted. This necessarily means that aC ¼ aB ¼ aA ¼ 0 (i.e., all the
a associated with the blocks above D) according to constraint (7). On the other hand,
constraint (6) for the blocks above D are shown as follows:
xC þ aC  xB ; ð11Þ

x B þ aB  x A : ð12Þ

If block C is not extracted, the left side of restriction (11) equals 0, forcing xB to be 0
as well. In consequence, the left side of restriction (12) is also 0, forcing xA to be 0
as well. The propagation of this restriction forces all the blocks above D to not be
extracted in this example.
Both restrictions combined avoid the non-convexity issues presented before,
ensuring a better stope shape. However, the addition of these new constraints with the
new variable to the original problem generates the loss of the network structure. This
structure is based on the fact that the restriction set must have only precedence
constraints in the form of restriction (5), with only two variables for each restriction.
The new constraints do not have this structure, therefore, this problem can not be solved
using maximum flow algorithms. Branch and bound algorithms will be used instead, in
order to find the optimum stope design. These algorithms are slower than maximum
flow algorithms, hence the runtimes of this new problem will be studied as well.

3 Results

3.1 No convexity constraints

First of all we present the results obtained from the integer program implementation
(IP-formulation) without convexity constraints solved with simplex and branch and
bound algorithm implemented in Gurobi 5.6.3. This problem is equivalent to the

123
 G. Nelis et al.

network problem solved by Bai with the maximum flow algorithm. Therefore, in
order to validate the new implementation, we will compare the performance of both
algorithms solving the same instance. A real orebody, presented in Fig. 7, will be
used with the economical and geomechanical parameters shown in Table 1. The
orebody is located in Canada [commodity (a metal) and exact location are
undisclosed for confidentiality reasons]. It has been the object of a resource estimate
report (NI-43-101) and could eventually be exploited by sublevel stoping.

Also, with the aim to test the performance in bigger instances, the cylindrical
block model discretization Dr=Dx was changed, where Dx is the original block size
in the x direction. Changing this value generates different instances with different
sizes as the number of blocks in the discretized cylindrical system equals the
number of variables for the instance without convexity constraints. The performance
of both algorithms is presented in Table 2.
As shown in Table 2, both methods achieve the same results in every field. Note
that the small differences between the objective functions (profits in the cylindrical
system) and the stope profits (profits in the Cartesian system) are due to the back-
transformation to the Cartesian coordinates required for mining. The stope shape is
also the same in both algorithms, as it can be seen in Fig. 8. With these results, we
validate the IP implementation of Bai’s methodology.
On the other hand, if we compare the runtime of the two algorithms, the maximum
flow implementation solves the same instances up to four times faster than Gurobi.
However, for the integer nature of this formulation, the results obtained by Gurobi
are notably fast. This is due to the structure of these linear IPs: the restriction matrix
generated by all the precedence constraints is totally unimodular. Therefore, the
linear relaxation of these instances, solved by simplex, generates an integer solution
without using the branch and bound techniques. This property causes a runtime

Fig. 7 Real deposit used in this analysis

123
Stope optimization with vertical convexity constraints

Table 1 Geomechanical, design and economic parameters for stope optimization algorithm
Parameters Real deposit Deposit 1 Deposit 2

Design and geomechanical parameters

Minimum stope width (m) 6 6 6
Minimum footwall angle ( ) 60 63 60
Minimum hangingwall angle ( ) 45 63 45
Economic parameters
Unitary cost ($/t) 50 50 50
Cut-off grade 0.005 0.005 0.005
Selling price ($/kg) 10 10 10
Recovery rate (%) 90 90 90
Density (t/m3 ) 2.3 2.3 2.3

Table 2 Comparison between network structure: flow algorithm and IP formulation—Gurobi simplex
algorithm
Dr=Dx Algorithms Objective function Stope profit Missed ore Dilution Runtime
(k$) (k$) (k$) (%) (s)

1 Flow 1184.27 1202.24 179.84 4.83 23.15

Gurobi 1184.27 1202.24 179.84 4.83 69.87
0.5 Flow 1150.78 1196.60 187.68 3.57 43.06
Gurobi 1150.78 1196.60 187.68 3.57 148.43
0.3 Flow 1145.03 1199.63 184.16 3.68 97.78
Gurobi Out of memory error

Fig. 8 Vertical sections (Y = -68) of optimal stopes generated by maximum flow algorithm and Gurobi
simplex algorithm after back-transformation to Cartesian coordinates

remarkably low compared to other instances with comparable size. Thereby, the IP-
formulation is still useful for solving these instances.

3.2 Convexity constraints

To study the effect of the convexity constraints proposed before, three different
orebodies were used. The first one corresponds to the real deposit used in the

123
 G. Nelis et al.

Table 3 Stope design results of IP-formulation for real deposit

Dr=Dx Cases Stope profit (k$) Missed ore (k$) Dilution (%)

1 No convexity 1202.24 179.84 4.83

Convexity (gap 0.1 %) 1200.86 180.15 5.39
Convexity (gap 0.01 %) 1200.87 181.08 4.93
0.9 No convexity 1200.73 181.60 4.52
Convexity (gap 0.1 %) 1198.72 182.57 5.14
Convexity (gap 0.01 %) 1199.70 182.49 4.57
0.8 No convexity 1195.73 186.82 4.42
Convexity (gap 0.1 %) 1194.41 187.39 4.84
Convexity (gap 0.01 %) 1194.53 188.05 4.45

No convexity IP-formulation without convexity constraints, Convexity (gap 0.1 %) IP-formulation with
convexity constraints. Minimum optimality gap: 0.1 %, Convexity (gap 0.01%) IP-formulation with
convexity constraints. Minimum optimality gap: 0.01 %

previous section, presented in Fig. 7, and its results with the convexity constraints
are presented in Table 3 along with the results obtained using Gurobi on the same
instance without the new constraints, to compare both results graphically and
quantitatively. In terms of value, for this orebody, the difference between both cases
is less than 1 %, with a small raise in the dilution inside the stope, and less ore
recovery. This behavior is expected due to the fact that the new constraints reduce
the feasible region of the initial problem.
In relation to runtimes of both algorithms, the addition of the new constraints
generates a significant increase of the solving time for the same deposit. This is due
to the fact that the restriction matrix is not totally unimodular under the new
constraints, therefore the simplex solution is not integer and branch and bound
techniques must be applied, with the notable increase in the delivering time of the
optimum stope, being up to 20 times slower than the original formulation. In this
regard, it is pertinent to emphasize the magnitude of these instances in terms of
variables and constraints (see Table 4). Compared to the original formulation, the
number of variables is doubled as it is expected, due to the new a variable. Also, the
number of constraints increases up to 3.2 times, generating an instance of nearly 3
millions of constraints with the smallest discretization presented here.
For the synthetic deposits, two different cases are shown. The first deposit
represents an adverse case for the original algorithm since it contains high value on
top and bottom zones, and a narrow band in its center (Fig. 9a). This geometry
generates a non-convex zone in its midsection. The second case, shown in Fig. 9b,
represents a deposit with scattered high grade blocks through all the orebody,
producing irregular contours. The results for both orebodies are presented in Table
5.
For both cases, the addition of the convexity constraints lowers the stope profit,
which is a expected result. This effect is different for each deposit: The first deposit
presents the more adverse case for the convexity constraints, with a loss of 14 % in
the stope profit and a dilution increase from 4.49 to 14.99 % for the smallest

123
Stope optimization with vertical convexity constraints

Table 4 Optimization results of IP-formulation for real deposit

Dr=Dx Cases Objective Optimality Constraints Variables Runtime (s)
function (k$) gap (%)

1 No convexity 1041.60 – 484,390 73,944 3.7

Convexity 0.1 1040.47 0.0631 1,337,590 147,888 45.4
Convexity 0.01 1040.95 0.0096 1,337,590 147,888 47.7
0.9 No convexity 1070.14 – 669,578 98,000 5.5
Convexity 0.1 1069.10 0.0511 1,992,578 196,000 128.6
Convexity 0.01 1069.28 0.0080 1,992,578 196,000 154.0
0.8 No convexity 1056.75 0 920,033 138,688 8.1
Convexity 0.1 1055.73 0.0522 2,935,343 277,376 229.4
Convexity 0.01 1056.15 0.0071 2,935,343 277,376 280.4

No convexity IP-formulation without convexity constraints, Convexity 0.1 IP-formulation with convexity
constraints. Minimum optimality gap: 0.1 %, Convexity 0.01 IP-formulation with convexity constraints.
Minimum optimality gap: 0.01 %

Fig. 9 Isometric views of both synthetic deposits used in this analysis. Only blocks with positive profit
value are shown

Table 5 Stope design results of IP-formulation for synthetic deposits

Dr=Dx Cases Stope profit (k$) Missed ore (k$) Dilution (%)

Deposit 1
1.5 No convexity 340.77 42.43 8.41
Convexity (gap 0.5 %) 305.09 40.39 17.77
1 No convexity 355.26 41.75 4.49
Convexity (gap 0.5 %) 306.39 52.43 14.99
Deposit 2
1 No convexity 984.27 33.14 23.01
Convexity (gap 0.5 %) 947.33 47.50 30.23
0.9 No convexity 993.40 17.18 24.48
Convexity (gap 0.5 %) 958.13 31.56 30.79

123
 G. Nelis et al.

Table 6 Optimization results of IP-formulation for synthetic deposits

Dr=Dx Cases Objective Optimality Constraints Variables Runtime (s)
function (k$) gap (%)

Deposit 1
1.5 No convexity 373.78 – 20,600 3920 0.94
Convexity 0.5 324.97 0.47 34,712 7840 2.245
1.0 No convexity 359.41 – 64,308 11,760 2.86
Convexity 0.5 308.13 0.48 129,828 23,520 166.77
Deposit 2
1 No convexity 949.80 – 142,608 21,840 6.17
Convexity 0.5 923.11 0.48 394,608 43,680 109.60
0.9 No convexity 965.85 – 195,118 28,644 7.42
Convexity 0.5 939.08 0.49 581,812 57,288 265.68

discretization (Dr=Dx ¼ 1). The dilution increases since the solution must add new
waste blocks to meet the new constraints. Regarding deposit 2, the profit loss is
around 4 % with a small increase in dilution.
Moreover, it is relevant to mention that the addition of convexity constraints
produces a greater amount of ore missed compared to the original solution in almost
every instance. This can be interpreted as a trade-off between ore and value: in some
zones of the deposit it is better to leave certain ore blocks out of the solution in order
to meet the convexity constraints. Adding these extra ore blocks would add too
much dilution to the solution.
The differences between the resolution time with and without convexity
constraints vary greatly depending on the size of the problem (see Table 6). For
the first deposit, for instance, with a discretization of Dr=Dx ¼ 1:5; the runtime is
almost tripled when convexity constraints are used. On the other hand, with a
discretization of Dr=Dx ¼ 1:0; the solving time is up to 58 times greater for the
same deposit. This indicates the great impact of the instance size on the solving
times. For example, despite being the same deposit, increasing the variables and
constraints using a smaller discretization generates a solving time 2.5 times greater
for the lowest gap shown in Table 6 for deposit 2: from 109.6 to 265.68 s.

3.3 Stope shape and stability

To study the stope stability on the new solutions, we compare different stope
designs in Examine2D, showing the SF using Hoek and Brown (Hoek et al. 2002)
and Mohr–Coulomb (Coulomb 1776) failure criteria according to the parameters
and geomechanical conditions shown in Table 7. The results for the real deposit
(Fig. 10) are presented in Fig. 11. Those for the synthetic deposit (Fig. 12) are
presented in Figs. 13 and 14.
In Fig. 11, the original stope presented a non-convex zone in its corner,
generating a failure by tension zone. The new stope, on the other hand, does not
present that tension zone, achieving a greater stability in that corner. Furthermore,

123
Stope optimization with vertical convexity constraints

Table 7 Geomechanical and

Parameters Values
failure criterion parameters
In situ stress parameters
Depth (m) 450
Overburden unit weight (MN/m3 ) 0.026
Horizontal stress ratio 1.5
Out of plane stress ratio 1.2
Rock mass strength parameters
Tensile strength (MPa) 0.3
Cohesion (MPa) 2
Friction angle ( ) 50
High quality rock
Intact compact strength (MPa) 250
GSI 75
mi 25
Low quality rock
Intact compact strength (MPa) 100
GSI 30
mi 25

Fig. 10 Stope shapes of both algorithms applied on real deposit

123
 G. Nelis et al.

Fig. 11 Stability analysis of a cross-section on real deposit stope

Fig. 12 Stope shapes of both algorithms applied on deposit 1

the stope stability in other zones also improves as we can see in the upper-left
corner, where the original stope had an unstable zone that was improved
significantly with the new constraints. The induced stress over this zone changes
due to the regularization of the stope. Therefore, the convexity constraints can
improve not only the local stability, but the global stability as well.
Similar observations can be made for the synthetic deposit (Fig. 13). The failure
zone in the mid-zone is substantially reduced with the new convexity constraints.

123
Stope optimization with vertical convexity constraints

Fig. 13 Stability analysis of a cross-section on synthetic deposit 1. High quality rock mass

Fig. 14 Stability analysis of a cross-section on synthetic deposit 1. Low quality rock mass

This behavior holds true for different rock mass qualities, as it can be seen in
Fig. 14, where we evaluated the same stope geometry, under the same stress
conditions, but with a low quality rock mass detailed in Table 7. Therefore, the new
geometries generated by the convexity constraints are more stable than the stopes
from the original methodology, independently of the rock mass quality.

4 Discussion

The new IP-formulation presented in this work improves the stope designs obtained
by the network formulation by Bai et al. (2013a). The addition of the convexity
constraints achieves more regular and geomechanically more stable shapes without

123
 G. Nelis et al.

using an additional step to manually regularize the stope borders with the risk of
loosing the design optimality.
In the examples presented, the stopes obtained under the new IP formulation are
comparable with the stopes obtained by Bai’s methodology, with differences that
range from less than 1 to 13 % in the stope profit function. However, depending on
the cost structure, the stope profit reduction could be greater with the IP formulation,
since it forces to add waste blocks or to eliminate ore blocks from the original
solution to fulfill the convexity constraints. These may be critical in low-grade
deposits where a subtle change in the stope design may produce a great change in its
profit.
It can be noticed, also, that the new constraints structure allows the solution to
add or discard blocks from the original solution obtained by Bai’s methodology,
depending of the value generated by these actions. This is a clear advantage over a
manual adjustment stage since it is not trivial to determine whether to add or
subtract blocks to achieve a better geometry, while seeking the best possible value
and respecting the diagonal precedence constraints to maintain the footwall and
hangingwall angles.
Regarding stability, all the new designs are more stable than their previous
solutions obtained without the constraints. However, there are still some areas with
lower stability despite complying with all the restrictions imposed on the
formulation. This can be solved by generating smaller stopes, modifying the stope
height (controlled by the raise length in the original Bai’s algorithm) and the stope
length (controlled by the parameter R in the original Bai’s algorithm). These
parameters must be selected under strict geomechanical criteria given the stress
conditions and the rock mass quality where the stope will be built.
In terms of regularity of the stope design obtained by the IP formulation, we can
see a great improvement in all the instances shown compared to Bai’s algorithm.
This will ensure an easier work for the production and construction phase in-field.
However, the final stope design could still be modified in real life work given the
blasting control and the presence of joints or other structures.
The stope design could be further regularized using the same convexity
constraints structure, but with different coordinates. For instance, in the cylindrical
coordinate system, we could take rings instead of columns: the group of blocks with
same r and z coordinate, but different h: If we apply the same restrictions but
changing these coordinates, we will avoid the irregularities not only vertically, but
horizontally as well. This regularization will improve the stope geometry, but it will
worsen the stope profit, therefore, the decision to include these new constraints will
depend on the mining engineer. As a downside, adding these new constraints would
increase the variables and restrictions notably, and the instances shown are already
remarkably large.
Despite the low resolution times for these big instances, the definition of the
instance using Gurobi requires a large amount of RAM. This is a problem in some
cases, where the sole definition of variables and restrictions in Gurobi consumes all
the available RAM. This is not an issue with the network structure used by Bai et al.
(2013a) since it is more compact, requiring less memory to be defined and solved.
However, these big instances are not common in Sublevel Stoping given its

123
Stope optimization with vertical convexity constraints

selective nature. Admittedly, this issue could become a problem when multiple
stopes have to be optimized simultaneously as in Bai et al. (2014).
A possible alternative to reduce the size of the instance would be the
implementation of a heuristic that, instead of applying the convexity constraints
in every column of the block model, focus only in its outer columns, controlled by a
parameter q defined from the initial raise. From r [ q; this heuristic would apply
the convexity constraints, leaving the inner part of the block model without these
restrictions. This would take advantage of the fact that the inner parts of the block
model does not present convexity issues. This heuristic could achieve good results
faster than the IP-formulation and it would work with larger instances.
A usual issue in stope optimization is avoiding stope overlapping when multiple
stopes are considered. Bai’s formulation explicitly avoids this issue defining a single
raise for each stope, and a different cylindrical coordinate system for each one of
them, where each block of the original block model belongs only to one of these
systems, as it is shown in Bai et al. (2014). Since the convexity constraints are
applied to each one of these different coordinate systems, the IP-formulation does
not generate overlapping stopes.
Finally, regarding runtimes, the IP-formulation is slower than the network
structure solving the same instances, and the addition of convexity constraints
increases the runtimes even more. However, this formulation remains remarkably
fast for this kind of problems: the stope design definition problem belongs to the
strategic mine planning stage, where decisions takes months or years to be made, so
fast runtimes are not mandatory. For this reason, the IP-formulation with convexity
constraints could be used in real-life applications.

5 Conclusions

The proposed methodology allows the determination of optimal stope designs,

fulfilling geomechanical requirements. It generates feasible stopes in disseminated
orebodies, avoiding unstable and irregular geometries. The designs comply with
technical requirements such as the stope drilling pattern. The formulation can be
solved in reasonable times for real life cases, and it can be extended to allow further
regularization of stope design or faster resolution times using heuristics.

Acknowledgments The Government of Canada provided a Research Scholarship to the first author. The
authors are grateful for the comments, corrections and insights made by the three anonymous reviewers.

References
Alford C (1996) Optimisation in underground mine design. In: Proceedings of 25th APCOM. AusIMM,
Melbourne, pp 213–218
Ataee-Pour M (2000) A heuristic algorithm to optimise stope boundaries. PhD Thesis, University of
Wollongong
Ataee-Pour M (2005) A critical survey of the existing stope layout optimization techniques. J Min Sci
41:447–466

123
 G. Nelis et al.

Bai X, Marcotte D, Simon R (2013a) Underground stope optimization with network flow method.
Comput Geosci 52:361–371
Bai X, Marcotte D, Simon R (2013b) Incorporating drift in long-hole stope optimization using network
flow algorithm. In: Proceedings of 36th APCOM. APCOM, Porto Alegre, p 391–400
Bai X, Marcotte D, Simon R (2014) A heuristic sublevel stope optimizer with multiple raises. SAIMM
114:427–434
Bieniawski Z (1989) Engineering rock mass classifications: a complete manual for engineers and
geologists in mining, civil, and petroleum engineering. Wiley, New York
Cheimanoff N, Deliac E, Mallet J (1989) GEOCAD: an alternative CAD and artificial intelligence tool
that helps moving from geological resources to mineable reserves. In: 21st Application of computers
and operations research in the mineral industry: 21st international symposium: papers, p 471
Coulomb CA (1776) Essai sur une application des regles des maximis et minimis à quelquels problèmes
de statique relatifs a l’architecture. Mem Acad Roy Div Sav 7:343–387
Crouch S, Starfield A (1983) Boundary element methods in solid mechanics. George Allen and Unwin,
London
Hoek H, Carranza-Torres C, Corkun B (2002) Hoek–Brown failure criterion—2002 edition. In:
Proceedings of the fifth North American rock mechanics symposium (NARMS-TAC). University of
Toronto Press, Toronto, p 267–273
Lane K (1988) The economic definition of ore: cut-off grades in theory and practice. Mining Journal
Books, London
Laubscher DH (1990) A geomechanics classification system for the rating of rock mass in mine design.
J S Afr Inst Min Metall 90:257–273
Lerchs H, Grossmann I (1965) Optimum design of open-pit mines. Trans CIM 58:47–54
Manchuk J, Deutsch C (2008) Optimizing stope designs and sequences in underground mines. SME Trans
324:67–75
Mathews KE (1980) Prediction of stable excavation spans for mining at depths below 1000 m in hard
rock. Report to Canada Centre for Mining and Energy Technology (CANMET), Department of
Energy and Resources
Ovanic J, Young D (1995) Economic optimisation of stope geometry using separable programming with
special branch and bound techniques. In: Third Canadian conference on computer applications in the
mineral industry, p 129–135
Picard J (1976) Maximal closure of a graph and applications to combinatorial problems. Manag Sci
22:1268–1272
Riddle J (1977) A dynamic programming solution of a block-caving mine layout. In: Application of
computer methods in the mineral industry: proceedings of the fourteenth symposium, p 767–780
Sandanayake D, Topal E, Ali Asad M (2015) A heuristic approach to optimal design of an underground
mine stope layout. Appl Soft Comput 30:595–603
Topal E, Sens J (2010) A new algorithm for stope boundary optimization. J Coal Sci Eng 16(2):113–119

123