Professional Documents
Culture Documents
Articulo 22
Articulo 22
Articulo 22
To cite this article: Levente Sipeki, Alexandra M. Newman & Candace A. Yano (2019):
Selecting Support Pillars in Underground Mines with Ore Veins, IISE Transactions, DOI:
10.1080/24725854.2019.1699978
Article views: 12
t
ip
2Industrial Engineering and Operations Research Department and the Haas School of
cr
Business University of California at Berkeley
us
*Corresponding Author: Email: newman@mines.edu
an
Abstract
M
We address the design optimization problem for a mine in which the ore is concentrated
d
in a system of long, thin veins and for which the so-called top-down open-stope mining
e
method is customarily used. In such a mine, a large volume of earth below the surface
pt
are left behind as part of a pillar to provide geotechnical structural stability; the
remainder are extracted and processed to obtain ore. We seek a design that maximizes
Ac
t
Constraints, Integer Programming
ip
History: Submitted November 2018; accepted November 2019 after one revision.
cr
us
1 Introduction
Our work was motivated by mining contexts in which the ore is concentrated in a system
an
of very long, thin veins containing rock with high ore density, as is common for gold,
silver, lead, zinc, nickel, and iron ore. Unlike mineral deposits involving soft substances
M
such as coal, which are usually flat-lying, veins arise in hard rock and generally span a
substantial vertical distance because of how the deposits formed. For example, gold is
d
deposited in the rock by the intrusion of hydrothermal solutions. The ore can
e
nevertheless be extracted via the top-down open-stope mining method, under which the
pt
extraction progresses layer-by-layer in the downward direction, and on each layer in the
mine, the extraction process creates open stopes (areas of void) while unextracted
ce
material provides structural integrity. For such a setting, mining companies conceptually
divide the vein and the surrounding rock into three-dimensional rectangular blocks on
Ac
Figure 1 shows a schematic in which the left panel represents a vertical slice of earth.
There is a solid vertical “wall” shaded in light grey, depicting an area containing
negligible ore that is not economically worthwhile to extract. Between any two such
walls is a vertical, single-block-deep “brick wall” composed of blocks (shown in dark
grey) that may be extracted, depending upon their ore density. In order to extract ore
t
from underground mines, infrastructure must be built so as to allow mining equipment
ip
access to different depths of the ore body. Top-down open-stope mining allows
cr
extraction to begin, with minimal infrastructure development, at a level in the deposit
that is expected to be profitable. Then, rather than incurring the cost of backfilling
us
extracted areas as a way to maintain geotechnical stability, pillars, i.e., rock that already
exists in the mine, act as supports against the lateral earth pressures. The right panel of
an
Figure 1 shows how the face of a front-to-back slice through a mass like that shown on
the right-hand side in the left panel would appear for the case of two parallel veins. In
M
the diagram, on the left and right and in the middle, running from top to bottom, are solid
vertical walls. One or more consecutive blocks (forming a stope) are removed from one
d
horizontal layer of the “brick wall” that contains a vein; this process is repeated in other
e
locations, creating voids. Horizontal pillars are left in situ to preclude caving of the walls
pt
t
ip
We devise a new way to capture a significant portion of the remaining stability
considerations via an iterative constraint generation method utilizing “barely infeasible”
cr
voids, i.e., voids that would be evaluated as geotechnically stable if a single block in
us
each void had not been extracted. We developed this approach utilizing barely
infeasible voids specifically for our problem. We then demonstrate that our methodology
is computationally viable and performs well for a large, real problem, producing very
an
good solutions in about two hours of CPU time or less. A final evaluation of geotechnical
stability via time-consuming finite element analysis is necessary, but we have found that
M
systematic inclusion of relatively simple constraints is adequate for the mine design to
pass the final test. In a real case study, our procedure produced a solution that was
d
evaluated to be geotechnicaly stable and with an estimated profit 16% larger than the
e
The remainder of this paper is organized as follows: Section 2 contains a review of the
ce
2 Literature Review
We focus our literature review on commonly-utilized underground mining methods in
which stopes (open voids) are created, either temporarily or permanently, as part of the
extraction process. With such methods, it is necessary to ensure geotechnical stability
until the voids are backfilled, or permanently, depending upon the mining method.
t
entails backfilling voids resulting from extraction, so geotechnical stability is easier to
ip
maintain. Mining engineers need to determine which areas to extract and which to leave
cr
behind with the goal of maximizing profit, taking into account the concentration of ore in
various parts of the mine, cost of extraction, and constraints on mineable shapes.
us
Examples of research in which the authors seek to (in most cases, approximately)
optimize profit include Alford (1996), Alford et al. (2007), Ataee-Pour (2000, 2005), Bai
an
et al. (2013), Nelis et al. (2016) and Grieco and Dimitrakopoulos (2007).
M
Researchers have also studied geotechnical factors that affect how pillar sizes and
locations influence the stability of mines in which open voids are left behind (such as
d
those that arise in top-down open-stope mining), including Starfield Wawersik (1968),
e
Mathews et al. (1981), Hamrin and Hustrulid (2001), Tesarik et al. (2009), and Machuca
et al. (2015). There are also many articles on approaches that extend or modify the
pt
Mathews et al. methodology, among others. The aforementioned articles focus primarily
ce
on local stability issues and do not seek an optimal comprehensive mine design.
Zipf (2001) takes a more holistic view and considers both regular pillars and massive
Ac
barrier pillars that can limit the extent of cascading failures but focuses on stability
conditions and does not seek to optimize the mine design.
To the best of our knowledge, there is no research on optimizing the full design of mines
employing top-down, open-stope extraction in hard rock in which open voids are left
behind after excavation is complete. We focus on vein-system mines and take
advantage of their structure in developing an efficient and effective solution procedure.
3 Design of Top-Down Open-Stope Mines with a
Single Vein
In the introduction, we briefly described the topology of the types of mines that we
consider. Here, we first discuss the geometry of pillars and stopes and then elaborate
on geotechnical stability considerations and related mine design decisions. Throughout
the paper, we assume that the length is measured perpendicular to the principal
direction of stress.
t
ip
Recall that each vein is contained in a one-block-deep (from front to back) “brick” wall,
an example of which—-after some extraction—-is shown in the front-face of Figure 1.
cr
Due to characteristics of the block layout and the mining process, each stope consists
us
of a series of contiguous blocks on a single level of the mine. Each stope is extracted in
a continuous process, terminating when reaching a block that is designated to be left in
an
situ as part of a pillar. The process then restarts at the beginning of another stope. As
such, both pillars and stopes consist of contiguous blocks on a single level. Although
M
each individual continuous extraction operation is confined to a single level, as
extraction progresses downward from level to level, voids may become larger when new
d
stopes are extracted from lower levels. For example, in the left panel of Figure 1, the
e
large void in the lower right of the mass (labeled “Open Area #2”) extends from the
second-highest level down to the bottom level. For such a large void, there is
pt
considerable risk of the adjacent vertical walls failing due to lateral earth pressures.
ce
The allowable sizes and locations of pillars depend on (i) the rock strength and (ii) the
lateral earth pressures, which translate to lateral compressional stresses exerted on the
Ac
pillars. Ideally, structural stability should be evaluated using a finite element method
such as that available in industry-accepted software, e.g., AutoCAD Map3D (Autodesk
2019), but such methods require considerable CPU time to evaluate even a single
design and many alternative designs must be evaluated at the high-level planning
stage. Therefore, for pillars with (near-)rectangular shapes, engineers use a few
different industry-accepted proxy metrics to determine whether a given set of pillars can
withstand the stress exerted on them (Wagner 1980, Lunder and Pakalnis 1997). We
discuss each in turn.
An adequate length-to-width ratio of each pillar, which depends upon the strength of the
rock and the magnitude of the lateral pressure, ensures that each pillar itself is stable.
Due to the allowable form of a pillar and the fact that the blocks are roughly the same
size and shape, this constraint effectively translates to a lower limit on the number of
blocks in a pillar. This requirement may not be enough, however. For each potential
t
pillar, mining engineers define a tributary area that the pillar is expected to support and
ip
impose a constraint on the extraction ratio, which is the fraction of the tributary area that
cr
is actually extracted. An example of a tributary area is shown as a hexagon in the left
panel of Figure 1. Pillars with larger length-to-width ratios (assuming the height is that of
us
a level in the mine) are stronger per unit volume, allowing for a proportionally greater
extraction percentage. Therefore, the length-to-width and extraction ratio constraints
an
need to be specified jointly.
M
Engineers also limit the size of large extracted voids without proximate support, such as
the one in the left panel of Figure 1 mentioned earlier, by imposing a constraint on the
d
void value, a measure of the “size” of each open area. Qualitatively, the constraint
e
ensures that the surrounding pillars are not too far apart from each other. If the shape of
the void (stope) is regular, the hydraulic radius, which is defined as the ratio of the area
pt
to the perimeter, provides an adequate proxy for the void value and is also relatively
ce
easy to calculate. For irregular stope shapes, however, whose boundaries must be
defined by mining engineers for the purpose of estimating geotechnical stability
Ac
because the stopes are not fully enclosed, the hydraulic radius becomes a less accurate
measure of geotechnical stability as the aspect ratio of a void increases (Milne et al.
1996). Therefore, Milne et al. (1996) introduce the radius factor:
1 1 n 1
f ·[ ]1 , (1)
2 n 1 r
which is the harmonic average of the lengths of n radii, measured from the center point
of an open area to the nearest supporting pillar, where the angles of the radii are
equally spaced around a 360-degree circle. An example of a void and the associated
radii are shown in the left panel of Figure 2. In the formula, r is the length of the radius
360i
at angular increment θ, with θ assuming the values n degrees, i 1,, n . (The scale
of the hydraulic radius and the radius factor are not the same, but with both, a smaller
value of the metric indicates that the void is more stable and the relationship between
t
them is well-understood for enclosed voids with regular shapes.)
ip
cr
Fortunately, Milne et al. (1996) have also shown that for rectangular shapes, if the
aspect ratio of the void is between 0.25 and 4.00, the hydraulic radius is an adequate
us
proxy for the radius factor. We take advantage of this result and utilize the hydraulic
radius as our (in)stability metric. Throughout the remainder of the paper, we use the
an
term void value to refer to the hydraulic radius or an approximation of it.
M
There is, however, a significant challenge in using either the hydraulic radius or the
radius factor metric for situations in which there are voids that are not enclosed except
at the outer boundaries of the mine and/or when the voids are punctuated by pillars.
d
The radius factor affords more accuracy than the hydraulic radius with individual, well
e
defined voids. However, its evaluation depends upon identification of the center point
pt
that yields the largest value of the radius factor, as it is that value that quantifies the
ce
instability of the void. (Finding this center point is a nonlinear optimization problem in
two variables representing the coordinates of the center point in a plane.) The left and
Ac
right panels of Figure 2 show, respectively, the radii for an arbitrary center point
depicted as a white dot, and an arbitrary boundary, displayed as a dotted line, of an
area that might be used as the basis for estimating the hydraulic radius of the actual
void.
As part of our algorithm, we have developed a way to avoid the complications described
above. In particular, in our methodology, it is necessary to compute the hydraulic radius
for only a small number of voids offline in advance. We provide details later in the
paper.
t
a simple rule can be used, namely, that the dependent pillars must be aligned in the
ip
principal direction of horizontal stress (Brady and Brown 2013). When this rule is
cr
combined with constraints on the proxy metrics discussed earlier, these rules suffice for
capturing the first-order geotechnical stability considerations for the purpose of
us
estimating the profitability of the deposit in the context of long-term planning. In Figure
3, we list the proxy metrics and describe the role of each. We also illustrate some
an
consequences of violating the various stability constraints: pillar buckling and spalling,
and curvature of stope walls that may lead to their failure.
M
We note that vein-system mines typically include only a few vein intersections within
d
any phase of excavation (an entire mine in the case of small mines or a single phase for
e
Our industry partner found that even the process of organizing the data to evaluate the
geotechnical stability of a single mine design is time-consuming, requiring
approximately 2.5 person-months of an experienced mining engineer’s time to represent
each proposed mine design, which includes defining entities in the ore body along with
the material properties of each, as well as a mesh that captures the interactions among
the entities. Both the entities and the mesh differ from one proposed design to another.
This is one reason why it is so important to identify a few designs with high anticipated
profit that are also highly likely to be geotechnically stable. Geotechnical stability of a
mine is extremely important so it is not surprising that considerable resources are
dedicated to its evaluation, both by the academic community and by mining firms, but
progress has been much slower in developing structured, formal optimization
approaches for mine design that can identify good candidate solutions within a
reasonable amount of CPU time.
Mining engineers seldom have access to, or the expertise to use, optimization software
t
to address the pillar selection problem. Even if they did, existing methods would not be
ip
able to handle problems of realistic size. In practice, one of two general strategies is
cr
used. The first one focuses on choosing the highest value blocks for extraction,
addressing geotechnical requirements secondarily. The other begins with extraction
us
patterns that are known to satisfy the geotechnical constraints and seeks to find the
most profitable option within a set. an
In the first, essentially greedy, approach, engineers identify some number of most
M
profitable blocks, and, assuming they are extracted, then determine pillar designs locally
to support the extracted areas. Given enough person-hours for trial-and-error analysis,
d
this approach can result in a near-optimal solution; however, due to time limitations,
e
engineers cannot consider all design variations and geotechnical constraints. As such,
this approach usually results in structurally unstable configurations whose profit
pt
In the second approach, which is more conservative, engineers: (i) generate one or
Ac
t
ip
above that has been divided into three dimensional rectangular blocks, which blocks
should be left behind as pillars, satisfying geotechnical structural stability constraints,
cr
and which should be extracted and processed, with the goal of maximizing total profit
us
from the extracted ore. The four main types of constraints are: (i) limits on the void
value; (ii) restrictions on the length-to-width ratios of the pillars; (iii) an upper limit on the
extraction ratio in each pertinent neighborhood; and (iv) constraints that link decisions in
an
proximate veins. In this section, we first explain how we conceptualize the problem and
the underlying decisions. We then present two formulations: (i) a formulation intuitive to
M
mining engineers that includes an explicit decision for each block; and (ii) a set
partitioning approach in which block decisions are implicitly considered in the
d
To simplify the representation of the deposit via a block model, mining engineers first
determine appropriate height and length dimensions for a block that offers an
Ac
acceptable granularity for scheduling extraction activities (Alford et al. 2007). Then,
geological statisticians estimate the profit obtainable from each block based on the
grade distribution of the ore. The top left portion of Figure 4 depicts the top view of an
example deposit and the corresponding block model, and the bottom left portion shows
the ore density, with darker shades indicating higher density. The right panels of the
figure display the corresponding front views of the same deposit, with the lower right
panel showing a possible pillar configuration. The blocks have a length of 5 meters (as
displayed on the x-axis), a height of 25 meters (as displayed on the z-axis), and a width
that depends on the cross-sectional size of the deposit (as displayed on the y-axis).
We assume that the block model is given and view the problem as one of classifying
each block as part of either a pillar or a stope. We account for two tiers of structural
stability: (i) at the individual pillar level, and (ii) for a collection of pillars. To address (i),
engineers ensure that a series of contiguous blocks on a single level (forming a pillar) is
structurally stable by imposing a lower limit on the length-to-width ratio, with the width
t
corresponding to the maximum width of all blocks that comprise the pillar. In practice,
ip
the widths of blocks may differ somewhat, and our methodology can account for most of
cr
this variation in a straightforward fashion. We explain the necessary modifications for
both straightforward and special cases in Section 4.3. To address (ii), engineers impose
us
two types of constraints for pertinent neighborhoods throughout the mine: (a) extraction
ratio and (b) void value. The extraction ratio constraints are enforced on defined
an
neighborhoods throughout the mine; the threshold represents the fraction among the
total required to be designated as pillars. In our formulation, this fractional value is
M
translated into the minimum number of blocks classified as part of a pillar. Because
adequate strength of individual pillars is enforced by the length-to-width ratio, the
d
extraction ratio ensures that the pillars collectively can withstand compressional
e
stresses (acting on the xz-plane). We assume that the strength of a pillar depends only
pt
on the aggregate strength of the contiguous set of blocks within it on a given level, and
not on the classification of blocks on other levels.
ce
There is also a lower limit on the size of a stope due to factors such as the space
Ac
required to maneuver equipment and the cost and time required to reposition equipment
to start a new stope. These are straightforward to consider in our formulation.
If there is a stope for which the peripheral pillars are too far apart, the resulting void may
be larger than allowed by geotechnical standards. In §9.6 and §13.4 of their book,
Brady and Brown (2013) provide evidence that suggests geotechnical stability is
assured in a practical sense if: (i) each vein is one block deep, (ii) each void is
supported by (at least) two opposing walls, and (iii) each open area is no larger than
void value considerations allow. Also see, for example, Mathews et al. (1981),
Potvin (1989) and Barton et al. (1974) for related discussion. To ensure that the
geotechnical constraints are satisfied, we would need to identify all sets of contiguous
blocks in a single vein in the x- and (or) z-plane such that if all blocks were extracted,
the resulting void value would exceed a threshold, and, for each set, impose a
constraint that requires at least one block to remain in situ. However, there is an
exponential number of such constraints. Moreover, with voids of differing shapes and
t
ip
sizes, this process becomes more difficult due to the computational effort required to
calculate the void value. We develop a methodology that allows us to avoid calculating
cr
void values except those for a small set of “barely infeasible” trial voids. Because we
us
only need to calculate a small number of void values, it is possible to handle irregular
void shapes. Implementation details appear in Section 5.
an
We next present two integer-programming formulations, the latter of which and its
associated solution methodology mitigate the computational challenges of solving
M
realistic instances.
d
is extracted or left in situ, respectively, and (ii) auxiliary variables correspond to the
retention of pillars and extraction of stopes to ensure that the combination of pillars is
Ac
sufficient to maintain geotechnical stability. The set consists of pillars that meet or
exceed the minimum required length-to-width ratio, and the set contains stopes that
meet or exceed the minimum length. These sets contain (not necessarily exhaustive)
collections of contiguous blocks; larger pillars and stopes can be constituted from sets
of adjacent smaller pillars or stopes, respectively. We generate sets of pillars and
stopes for inclusion in the optimization problem that collectively enable us to constitute
all feasible pillars and stopes. Set contains sets of contiguous blocks for which
extraction ratio constraints must be satisfied, while set contains sets of blocks, each
of which represents a “barely infeasible” void, i.e., one whose void value would be
feasible if one fewer block in it were extracted. Note that any voids corresponding to
sets that are supersets of those in are also infeasible, but analogous constraints
associated with these larger voids would be redundant.
In general, we use lower-case letters for parameters and upper-case letters for
variables. However, we also use lower-case letters for indices and upper-case script
t
letters for sets. A formulation follows.
ip
Indices and Sets
cr
: set of blocks {1,,|
us
|}
ˆ
e : set of blocks that form contiguous set e
ce
p
: set of blocks contained in pillar p
ˆ
b : set of stopes that contain block b
ˆ
b : set of pillars that contain block b
Parameters
nˆe : number of pillar-classified blocks required to satisfy the extraction ratio constraint in
set e
t
ip
Decision Variables
cr
1 if pillar p is left in situ
Xp :
0 otherwise
us
1 if stope s is extracted
Ys :
0 otherwise
an
1 if block b is left in situ
Wb :
0 otherwise
1 if block b
M
is extracted
Zb :
0 otherwise
d
A formulation follows.
e
g ·Z (2)
pt
(P) max b b
b
ce
s.t. W b nˆ e e (3)
b ˆe
Ac
W
b
b 1 v (4)
v
Ys Zb s , b s (5)
X p Wb p , b p (6)
Zb Wb 1 b (7)
X p Wb b (8)
p ˆb
Y s Zb b (9)
s ˆb
Due to the complementary nature of W and Z, we can trivially substitute one set of
variables out of the formulation, but we retain both for ease of presentation. The
t
ip
objective (2) is to maximize the profit realized from blocks that are extracted and
processed. Constraints (3) ensure that any set of blocks for which there is an extraction
cr
ratio constraint contains an appropriate number of blocks left in situ. Constraints (4)
us
ensure that for each set of blocks in which the void value constraint would be violated if
all blocks were extracted, at least one block is left in situ. Constraints (5) imply that if a
an
stope is extracted, all blocks contained in the stope must be extracted as well.
Constraints (6) preclude a pillar from remaining in situ without all its corresponding
M
blocks being left in situ as well. Constraints (7) ensure that a block is either extracted or
left in situ, but not both. Constraints (8) require that if a block is left in situ, (at least) one
pillar containing it is left in situ. Constraints (9) ensure that if a block is extracted, (at
d
We note that it is not possible to enumerate all sets for constraints (4). Furthermore,
even if we could enumerate them, adding all of the associated constraints to the
ce
…, ω can assume any integer value from n up to the maximum number of blocks in a
pillar. Thus, there is usually no need to include larger pillars in the formulation, as they
t
ip
are accounted for implicitly. Our approach keeps the number of pillars in the
optimization problem to the bare minimum, and for this reason, the process of
cr
generating pillars requires very little CPU time.
us
If block widths vary widely, it may not be adequate to express the minimum pillar size in
terms of number of blocks: pillars containing wider blocks would have larger minimum
an
pillar lengths. It is easy to account for this when generating feasible pillars for inclusion
in the formulation. However, our strategy for limiting the range of pillar lengths when
M
generating the set of pillars may not be adequate if block widths vary widely. We
illustrate this point via a simple example. Suppose one pillar contains at least one wider-
d
than-average block and a pillar adjacent to it contains only average-width blocks. If both
e
pillars are selected in the solution, then they would actually comprise a single pillar and
pt
the pillar length requirement would depend on the width of the wider-than-average
block. As such, the total length of the two adjacent pillars may be insufficient because
ce
the length of the pillar containing only average-width blocks does not account for the
wider blocks in the other pillar. On the other hand, the combined pillar is much longer
Ac
than the individual pillar that contained the wider-than-average block, so it may well be
long enough to satisfy the length-to-(maximum)-width requirement. If very large
disparities in block widths exist among proximate blocks, then pillars containing more
than 2n 1 blocks may need to be considered and constraints may need to be imposed
to disallow simultaneous selection of adjacent pillars which, as a single integrated pillar,
would not satisfy the length-to-(maximum)-width requirement. For ease of exposition, in
the remainder of the paper, we discuss (minimum) pillar sizes (lengths) in terms of
number of blocks, but our approach can be generalized to handle most realistic settings.
t
ip
We continue to use the same notation as before and introduce the following new
notation:
cr
Sets
us
e : set of pillars that contain at least one block in contiguous set e
an
v : set of pillars that contain at least one block in contiguous set v
M
Parameters
d
cˆ pe
: number of block(s) contained in pillar p that are also contained in the extraction
ce
ratio set e
(P) max g ·Y
s
s s (11)
s.t. cˆ pe ·X p nˆ e e (12)
p ˆe
X p 1v (13)
p v
X p Ys 1 b (14)
p ˆb s ˆb
X p , Ys binary p , s (15)
The objective (11) is to maximize the profit realized from extracted stopes. Constraints
(12) and (13) ensure that an appropriate combination of pillars is left in situ such that
their aggregate contributions to satisfying the extraction ratio and void value constraints,
cˆ
respectively, are sufficient. The values pe in (12) are parameters that engineers do not
t
ip
normally derive because of the amount of geological data and the significant
computational effort required to do so; therefore, most engineers typically perform more
cr
rudimentary analysis to estimate the contribution of each block to the satisfaction of the
us
extraction ratio constraint in each pertinent region. Constraints (14) ensure that each
block is assigned to exactly one stope or pillar selected in the solution. As in ( P ), it is
impossible to enumerate the elements contained in
an , and we address this limitation in
the same manner as mentioned above.
M
4.4 Multiple Vein Deposits
d
Figure 5 shows a top view of a multi-vein system for a single level, with veins outlined
by dashed black lines. In a block model for a vein network, veins are represented as
e
sets of contiguous blocks in the xy-plane whose proximity to each other in the z-plane is
pt
For cases in which two veins are close enough together that the extraction of blocks in
Ac
the two veins needs to be coordinated due to stress interactions between the blocks, if a
pillar-classified block exists in one vein, the block with the same (x, y) coordinates in the
other vein must be classified as a pillar as well. To ensure stability throughout the vein
system, we identify pairs of blocks that are sufficiently close in the z-plane to require the
same classification and impose constraints to ensure these requirements are satisfied.
We refer to the set containing all such pairs as .
We use part of an actual block diagram to illustrate an example of vein
interdependence. The schematic in Figure 6 shows the top view of a single layer of
blocks in two veins that merge into one in the middle of the schematic. We show dotted
lines between pairs of blocks that are interdependent: if one block within the pair is part
of a pillar (stope), the other must have the same designation. Expressing these vein
interdependence constraints is straightforward:
X b X b (b, b ) . (16)
t
ip
Although the geotechnical requirement for the minimum pillar length (size) in any single
cr
vein remains the same, when veins are close together or intersect, either explicitly or
implicitly generating all feasible pillars becomes more difficult, as we need to consider
us
pillars with much more complex shapes. We discuss one way to accommodate vein
intersections and the potentially more complex pillar shapes. Unless the number of
an
blocks in the intersection is large enough to satisfy the minimum pillar length
requirement, we need to devise modifications to the formulation to ensure that this
M
constraint is satisfied for potential pillars that span the vein intersection.
pillars, which are sets of blocks that do not meet the minimum pillar size (length)
e
requirement when considered alone, but do constitute one when combined with the
pt
block(s) in the vein intersection, and (ii) addition of special constraints involving these
ce
pseudo-pillars to ensure that actual pillars containing the blocks in the vein intersections
satisfy minimum size (length) constraints. Below, we describe our approach for handling
Ac
vein intersections.
Two veins may intersect in the form of a “T” (a “Y” intersection can be handled
similarly), or a cross (or “X”). We designate the block(s) in the intersection of the veins
as a pseudo-pillar and define other necessary pseudo-pillars along with an additional
constraint that ensures that any pillar(s) containing the block(s) in the intersection
satisfy the minimum size (length) requirement. Below, we provide an explanation under
the assumption that the intersection contains exactly one block, but the principles apply
to intersections that span two or more consecutive blocks.
“T” Intersection:
1. Designate the dark block highlighted in the schematic on the left side of Figure 7 as a
pseudo-pillar.
t
ip
2. Create pseudo-pillars, each of which consists of a set of blocks that does not
constitute a pillar of adequate size or contiguity when considered alone, but when the
cr
dark block is added to the set, constitutes a pillar that satisfies the minimum size
us
(length) requirement. For example, if a pillar must contain a minimum of three adjacent
blocks, then the pseudo-pillars are: (i) the two blocks to the left of the dark block, (ii) the
an
two blocks to the right of the dark block, (iii) the pillar to the left of the dark block and the
pillar to the right of the dark block; (iv) the pillar to the left of the dark block and the two
pillars to the right of the dark block; (v) the two pillars to the left of the dark block and the
M
pillar to the right of the dark block; (vi) the two pillars to the left of the dark block and the
two pillars to the right of the dark block; and (vii) the two pillars below the dark block.
d
These seven pseudo-pillars are shown in the crosshatched areas in Figure 8, numbered
e
as indicated above.
pt
As a valid pillar (i.e., one satisfying the minimum size (length) requirement) can be
ce
formed by combining the dark block with a valid pillar adjacent to it, we also include in
the additional constraint all valid pillars containing n, n + 1, , 2n 2 blocks that are
Ac
located immediately adjacent to the dark block, where n is the smallest allowable
number of consecutive blocks in a pillar. (We do not need to include pillars with 2n 1
blocks because the addition of the dark block to a pillar with 2n 2 blocks leads to a
pillar with 2n 1 blocks, which is the largest pillar size that we need to consider using
our scheme in which larger pillars can be constituted from smaller pillars.) Thus, to
ensure that any pillar containing the dark block is feasible, we need add a constraint to
the formulation to ensure that if the dark block is selected as a (pseudo-)pillar, then at
least one of the following must also be selected:
• one of the aforementioned pseudo-pillars (in (i) through (vii)); or
If any of the above is selected, the dark block combined with the selected pseudo-pillar
or pillar satisfies the minimum size requirement for a pillar. (Some of these pseudo-
pillars and pillars overlap with one another, so not all can be selected simultaneously;
t
ip
related logic is handled via standard set partitioning constraints.)
cr
To explain the pseudo-pillars needed for a cross intersection, it is instructive to first
us
consider a single vein. After we explain the necessary pseudo-pillars for this simple
case, we describe the full set of pseudo-pillars.
Cross Intersection:
an
M
1. Designate the dark block highlighted on the right side of Figure 7 as a pseudo-pillar
and consider the vein in the vertical dimension.
d
2. Create pseudo-pillars, each consisting of a set of blocks which, if combined with the
e
dark block, would constitute a pillar that satisfies the minimum size (length) requirement
pt
for a pillar. These are shown in Figure 9 for the case of a minimum pillar size of three
ce
blocks.
3. The full set of pseudo-pillars then consists of: (a) the pseudo-pillars shown in Figure
Ac
9; (b) the pseudo-pillars shown in Figure 9 but all rotated 90 degrees clockwise or
counterclockwise (so the pillar would be contained in the other vein); (c) pseudo-pillars
consisting of all possible (i.e., 6×6) unions of the crosshatched blocks from the
pseudopillars in sets (a) and (b) above. We briefly describe one example. Consider the
pseudo-pillar comprising the union of the crosshatched blocks in (viii) and its 90-degree
rotation. Observe that the former pseudo-pillar contains crosshatched blocks
immediately above and below the dark block, and that the 90-degree rotation of it
contains crosshatched blocks immediately to the left and right of it. The union then
consists of the four blocks immediately adjacent to (above, below, to the right of and to
the left of) the dark block.
We note that some combinations of pseudo-pillars in set (c) may not be geotechnically
viable depending upon the angle of intersection between the veins.
We add a constraint to the formulation that ensures that if the dark block is selected as
a pseudo-pillar, then at least one of the following is also selected:
t
ip
• one of the aforementioned pseudo-pillars (in (a)-(c)); or
cr
• a valid pillar consisting of n, n + 1, , 2n 2 consecutive blocks located adjacent to
us
the dark block.
an
Note that in each case, the added constraint ensures that if the dark block is selected,
the pillar(s) containing it constitute a valid pillar in the pertinent vein(s). As one can see
M
from the discussion above, constructing the set of pseudo-pillars can be tedious and the
size of the set expands as the minimum pillar length increases, but in practical settings,
d
If two veins share two or more sequential blocks, the concepts described above can be
generalized to define appropriate pseudo-pillars and associated constraints.
Ac
We introduce additional notation here; we continue to use notation given in Section 4.3,
and the notational conventions there apply.
Indices and Sets
ˆ
: set of pseudo-pillars in vein intersections
p
: set of pillars and pseudo-pillars, one of which must be selected if pseudo-pillar
p ˆ is selected
: set of pairs of blocks ( b, b ) that must have the same classification
t
ip
The optimization problem becomes:
cr
(Pˆ ) max g ·Y
us
s s
s (11)
s.t. (12) - (14)
X p X p (b, b ) (17)
an
p ˆb
p ˆ
b
M
X p X p p ˆ (18)
p
p
d
X p , Ys binary p , s (19)
e
pt
Constraints (17) force vein-dependent blocks that share the same (x, y) coordinates to
ˆ
ce
have the same classification. Constraints (18) ensure that if any pseudo-pillar in set
is selected in the solution, then a corresponding (pseudo-)pillar must be chosen to effect
Ac
a feasible completion.
5 Solution Methodology
The optimization problems ( P ) and ( P ), as well as their multiple-vein extensions, are
theoretically NP-complete because they are set partitioning problems (see constraints
(7) and (14) in ( P ) and ( P ), respectively) with side constraints. Set partitioning is known
to be NP-complete (see Theorem 3.5 in Garey and Johnson (1979)), so ( P ) and ( P ) are
guaranteed to be at least as hard. Fortunately, if we ignore the void value constraints,
test problems of realistic size can be solved to within 5% of optimality in a few minutes
of CPU time. Our algorithm, which we call PillarSelector , exploits this by starting with
the solution to the relaxed problem without void value constraints, and iteratively solving
it, identifying violations of void value constraints by checking whether voids that are
known to violate the constraint appear in the relaxed solution. If so, constraints to
disallow those voids are added to the formulation.
t
We now discuss how we implement our approach for handling the void value
ip
constraints. The Brady and Brown (2013) analysis mentioned earlier limits the structure
cr
of the geotechnical constraints that must be considered, but in order to determine
structurally stable designs, we would need to identify all voids that violate the void value
us
threshold, which is computationally prohibitive. For this reason, we construct a small set
of “barely infeasible” trial voids with the property that the inclusion of a single additional
an
block as part of a pillar would reduce the void value below the designated feasibility
threshold. See Figure 10 for the four trial voids that we use in our computational
M
procedure. In the figure, HR denotes the hydraulic radius. We discuss these trial voids
in more detail later.
e d
Recall that Milne et al. (1996) found that for rectangular shapes, if the aspect ratio of the
void is between 0.25 and 4.00, the hydraulic radius adequately approximates the radius
pt
factor. Although two of the trial voids that we use are not rectangular so the aspect ratio
ce
rule-of-thumb cannot be applied to them directly, they do not deviate significantly from a
rectangular shape and their smallest rectangular enclosures have aspect ratios between
Ac
0.25 and 4.00, so we compute the hydraulic radius for each exactly, but acknowledge
that the computed value is not as accurate an approximation of the radius factor as it is
for the rectangular voids. Because we utilize a small number of trial voids and the voids
are rectilinear, it would not be difficult to compute the radius factor for each as the basis
for estimating geotechnical stability. However, because we are using the trial voids to
find voids that may be identical to, but are more likely to be supersets of, the trial voids,
it is not clear that great precision is needed in estimating the radius factor of each trial
void.
Although we have justified our approach to account for void value constraints for the
types of mines that we are studying, we believe that this basic idea can be utilized in
other contexts, as well. That is, one can identify voids that are barely infeasible by any
criteria that are pertinent to the specific mine or mining process and utilize them in the
same way as in our algorithm.
t
ip
For any solution to a relaxed version of a problem, we perform the following process to
determine whether constraints need to be added to the formulation. For each possible
cr
positioning of a trial void on the entire face of the “brick wall” associated with each vein,
us
we determine whether all blocks are stope-classified. If so, the void value constraint is
violated. We then repeat this process for other trial void shapes until all constraint
an
violations associated with all trial void shapes are identified. The violated constraints are
added to the formulation and the optimization problem is re-solved. This process is
M
repeated until a feasible solution is found for the most recent relaxation. We cannot
identify all possible “barely infeasible” voids (because of the exponentially large number
d
of potential voids), so we cannot guarantee that the final solution is feasible, but this is
also true of many of the ad hoc approaches used at the company that motivated our
e
study. For cases in which the finite element analysis determines one or more voids to
pt
be too large, constraints can be easily added to require at least one pillar-classified
ce
block within the void, after which the optimization problem can then be re-solved and
the finite element analysis rerun.
Ac
contain blocks in the trial void, populate v , and add the associated void value
constraint to .
Step 4: Re-solve the integer program, and go to Step 2.
Observe that the added constraints are retained throughout the iterative procedure, and
therefore, the relaxed problem becomes (weakly) tighter at each iteration.
t
6 Results
ip
We first report on a case study in which we utilized our algorithm to optimize the design
cr
of an underground mine in Africa whose block model contains 5,230 blocks. Our
us
industry partner (a large mining company) provided profit estimates for the blocks, and
the geotechnical thresholds.
an
Figure 10 depicts the four trial voids that we use: (i) a compressed cross, which is
intended to mimic a shape that is as close as possible to a diamond (given the block
M
dimensions in our problems) while remaining “barely infeasible”; (ii) a “barely infeasible”
rectangle with a length-to-height ratio as close as possible to 3:1; (iii) a “barely
d
infeasible” rectangle with a length-to-height ratio as close as possible to 1:3; and, (iv) a “
e
We utilized CPLEX to solve the integer programming problems within our iterative
ce
procedure on a Dell Power Edge R430 server with two 2.6 GHz Intel Xenon processors
under the Linux operating system. For our motivating example, our set partitioning
Ac
approach generated a solution that included none of the barely infeasible trial voids in
Figure 10 after two iterations with a run time of 38 minutes. Mining engineers at the
company applied finite element analysis to our mine design and confirmed its
geotechnical stability. The middle panel of Figure 11 displays our mine design
graphically for a section of the highest layer of a section of the mine.
We use as benchmarks two mine designs determined by a senior mine planner at the
company using the greedy and “pattern-shifting” approaches described in Section 3.1.
The solution (for the same section of the mine) generated via the greedy approach is
shown in the left panel of Figure 11. The solution contains 35 structurally unstable
pillars and violates 96 constraints—35 void value constraints and 61 vein
interdependence constraints. This is not surprising, however, as these constraints are
the most difficult to enforce using the combination of a manual process and computer-
aided-design software. Two areas in which vein interdependence constraints are
t
ip
violated are evident in the greedy solution. Near (a) and (b) in the left panel, there are
pillars on the left side of the diagram but not on the right, but the latter are necessary
cr
due to the vein interdependence constraints. In our approach, the strict enforcement of
us
these constraints results in pillars being retained in interdependent veins (from left to
right across the figure) near (c) and (d) in the middle panel. Although it is not evident
directly from the figure, void value constraints are violated because of the extracted
an
blocks shown in the figure and extracted blocks in the layers below.
M
The right panel of Figure 11 shows (again, for the same section of the mine) the best
solution from the more conservative, geotechnical stable “pattern-shifting” approach.
d
When using this approach, mining engineers define a common pillar size and shape to
e
be utilized throughout the mine. Because the pillars are sized to be geotechnically
pt
stable for the thickest vein, the pillars in other parts of the mine may be larger than
needed for geotechnical stability. An example of such a pillar is shown near (f) in the
ce
right panel of Figure 11. The effect of constraining the sizes and shapes of pillars and
their relative positions is significant; the total profit from our solution is 16% higher than
Ac
that from the best “pattern-shifting” solution. Minor manual adjustments to the latter may
recover part of this gap, but it is not easy to make these changes without violating the
void value or extraction ratio constraints.
Although we utilized the barely infeasible voids in Figure 10 to obtain the results
reported earlier, we were interested in understanding the value of considering additional
trial voids. We utilized 15 additional “barely infeasible” trial voids (see Figure 12) to
check for violation of void value constraints within the solution identified by our
procedure and did not find any violations. Although the shapes of these trial voids are
not much different than those of the four trial voids used in the main procedure, due to
the 1:5 aspect ratio of the blocks, it was difficult to construct trial voids that were
significantly different from being a simple rectangle and were just barely infeasible. We
also note that the hydraulic radius generally declines as a void becomes more
elongated or as the boundaries become more irregular, so voids with such shapes have
hydraulic radii that are not much larger than, and may even be smaller than, their
t
ip
largest inscribed square voids. As such, it is generally adequate to check for violations
of the void value constraint using trial voids that are roughly rectangular and relatively
cr
close to being square. For mines with other types of veins, the decision-maker could
us
design trial voids suited to the specific topography, emphasizing trial voids whose
shapes most closely resemble areas containing highly concentrated ore, i.e., those that
are most likely to violate the void value constraint.
an
We constructed three other examples that are based on or mimic actual data, but were
M
evaluated by our industrial collaborator under assumptions that do not align with our
model (e.g., a different mining method was employed); as such, we do not have
d
benchmark solutions for them. Our aim in solving these example problems was to
e
these problems are smaller than that in our case study, containing from 1,450 to 4,542
blocks. We were unable to find a feasible solution for the 4,542-block mine using the
ce
block-based formulation, but the set partitioning approach required less than 1,000
seconds of CPU time to find a solution with an optimality gap of 5% or less. The smaller
Ac
problems required less than 250 seconds of CPU time to achieve a similar optimality
gap using the set partitioning approach. We also solved variants of these problems in
which we perturbed the profits from the various blocks and varied the rock strength,
which has the effect of changing the minimum pillar size. Not surprisingly, as the profits
from the blocks become more similar and less correlated spatially, the problems
become more difficult to solve. Also, as the rock strength increases, the problems
become more difficult to solve as fewer of the blocks need to be specified as part of a
pillar, so there are more extraction options available. Nevertheless, even with
perturbations to make the problems more difficult, solutions within a 5% optimality gap
could be obtained within about two hours of CPU time. This suggests that our
methodology is efficient enough to be used as part of a “what-if” analysis, possibly
leading to a set of good solutions, allowing the decision-maker to choose one based on
difficult-to-quantify factors, such as logistics.
t
ip
We present an optimization algorithm for the pillar selection problem in a top-down,
open-stope mining operation, which entails determining which blocks to extract to
cr
maximize profit while ensuring that the blocks forming pillars that are left in situ satisfy
us
the necessary geotechnical constraints to ensure stability of the mine. Our algorithm
starts with a relaxation of so-called void value constraints that prohibit voids that are too
an
large (and thus unstable), and then progressively adds these constraints using a set of “
barely infeasible” trial voids to identify violations. As such, we solve a series of
M
progressively tighter relaxations of the problem. Because it is not possible to identify all
barely infeasible trial voids, the final problem that we solve remains a relaxation, but our
d
algorithm is able to produce a solution within 5% of optimality for the final relaxation in a
e
little over two hours, even for perturbations of the problem data to make the the problem
more challenging.
pt
ce
Our algorithm is efficient partly due to the manner in which we have formulated the
problem, which significantly limits the number of pillars and stopes and thus also the
computational effort for generating and selecting them, and partly due to our use of trial
Ac
voids to iteratively account for the void value constraints. Our idea of utilizing trial voids
of different shapes to iteratively account for violated void value constraints is new, and
enables us to avoid including very large numbers of side constraints–the vast majority of
which are redundant– in the formulation at the outset. The relaxed problems are fairly
compact and require only a modest amount of CPU time to solve.
For one real-world problem instance that was checked for stability by geotechnical
engineers using time-consuming finite element analysis, our procedure yielded a
feasible solution whose anticipated profit is tens of millions of dollars higher than the
best solution identified by a senior mining engineer in the company. Because our
procedure is computationally efficient, in practice, it would be possible to identify more
barely infeasible trial voids and to incorporate them into our algorithm to provide
stronger assurances that geotechnical stability constraints are satisfied before the
solution is checked using the finite element analysis, which is a necessary “gold
t
ip
standard” test. Other approaches in current use are much more ad hoc and either
satisfy geotechnical constraints but forgo a significant amount of profit, or achieve a
cr
higher profit but are geotechnically risky.
us
Due to the efficiency of our algorithm, it can be used in several ways. First, it can narrow
the set of designs with high estimated profit that need to be evaluated by excluding
an
those that violate many of the geotechnical constraints. This can provide significant
value by reducing the workloads of scarce mining engineers, allowing them to focus
M
their evaluation efforts on more promising designs. In addition, because our approach
requires only a modest amount of CPU time considering the scale and economic impact
d
of the decisions, mine planners can use our approach for sensitivity analysis to explore,
e
for example, the impact of various physical or design parameters on the best mine
pt
16% higher profit than the best solution identified by a senior mining engineer. This
translates into many tens of millions of dollars of improved profitability—due to the
capability of an optimization model to explicitly capture economic tradeoffs and a
multitude of interrelated constraints.
Acknowledgments
t
ip
The first author worked on this research project while he was a doctoral student at the
Colorado School of Mines. The authors acknowledge our industry partners who
cr
provided the data and explained their engineers’ decision rules, and funding from Alford
us
Systems. We also thank Professors Marcos Goycoolea from the University of Adolfo
Ibañéz, Andres Guerra from the Colorado School of Mines, and Andrea Brickey from
an
the South Dakota School of Mines for providing additional insights regarding model
formulation and exposition.
M
References
d
Barton N, Lien R, Lunde J (1974) Engineering classification of rock masses for the
design of tunnel support. Rock Mechanics 6(4):189–236.
Brady BH, Brown ET (2013) Rock Mechanics: For Underground Mining (Dordrecht, The
Netherlands: Springer Science & Business Media).
t
ip
Garey MR, Johnson DS (1979) Computers and Intractability (W.H. Freeman and Co.,
San Francisco, CA).
cr
Grieco N, Dimitrakopoulos R (2007) Managing grade risk in stope design optimisation:
us
probabilistic mathematical programming model and application in sublevel stoping.
Mining Technology 116(2):49–57. an
Hamrin H, Hustrulid W (2001) Underground mining methods and applications.
M
Underground Mining Methods: Engineering Fundamentals and International Case
Studies 3–14.
d
pillar optimisation at Granny Smith Mine. Underground Design Methods 2015, Perth,
Australia.
Ac
Potvin Y (1989) Design guidelines for open stope support. Canadian Institute of Mining
and Metallurgy Bulletin 82(926):53–62.
t
ip
Starfield A, Wawersik W (1968) Pillars as structural components in room-and-pillar
cr
mine design. The 10th US Symposium on Rock Mechanics, American Rock Mechanics
Association, Austin, TX.
us
Tesarik D, Seymour J, Yanske T (2009) Long-term stability of a backfilled room-and-
an
pillar test section at the Buick Mine, Missouri, USA. International Journal of Rock
Mechanics and Mining Sciences 46(7):1182–1196.
M
Villaescusa E (2014) Geotechnical Design for Sublevel Open Stoping (Taylor & Francis
d
Wagner H (1980) Pillar design in coal mines. Journal of the Southern African Institute
pt
Zipf K (2001) Toward pillar design to prevent collapse of room-and-pillar mines. 108th
Annual Exhibit and Meeting, Society for Mining, Metallurgy, and Exploration, Denver,
Ac
CO.
t
: set of all “barely infeasible” voids whose presence is to be checked for in the most
ip
recent relaxed solution
cr
: set of all “barely infeasible” trial voids in which all blocks are classified as stopes in
the most recent relaxed solution
us
v : set of blocks in trial void v
v
an
: pillars that contain at least one block in trial void v
while elimination of “barely infeasible” trial voids has not been confirmed do
for all v
M
do
if all blocks
b v are classified as stopes then
d
add v to set
end if
e
end for
pt
for all v do
ce
X p 1 v
add the following constraint to the formulation: p v
Ac
end for
re-solve the problem with the newly-added constraints
end while
t
ip
cr
Fig. 1 The left panel shows a three-dimensional schematic of a portion of the mine.
us
There is a stope in the lower middle as well as two other labeled open areas (voids).
The hexagonal area around the Example Pillar shows the tributary area that the pillar
an
might be expected to support. The diagram in the right panel shows a slice of a side
view, illustrating the “walls” between the veins, stopes excavated within two veins, and
M
the orientation of the lateral earth pressures against which the pillars must provide
geotechnical resistance. Further details appear in the main text.
e d
pt
ce
Ac
t
ip
cr
us
Fig. 2 Illustration of the data inputs required for the calculation of the hydraulic radius
an
and radius factor.
M
e d
pt
ce
Ac
t
ip
cr
us
an
M
e d
pt
ce
Ac
Fig. 5 Top view of a single level of a deposit, with black-dashed lines representing an
outline of the potential area to be extracted.
t
Fig. 6 Top-down view of two merging veins and interdependent blocks.
ip
cr
us
an
M
Fig. 7 T and cross intersections
e d
pt
ce
Ac
us
a cross intersection when minimum pillar size is three blocks
an
M
e d
pt
the colors outlining the blocks denote different ore density levels.
e
pt
ce
Ac
t
ip
cr
us
an
Fig. 12 Additional 15 shapes of “barely infeasible” voids tested before executing finite
element analysis. HR means hydraulic radius.
M
e d
pt
ce
Ac