Optimization of the billet cutting operation in the aluminum industry:

a case study

N. Brahimi1*, A. Khalaf2, H. Al-Hammadi3

1
Department of Industrial Engineering and Management, University of Sharjah,
PO Box 27272, Sharjah, United Arab Emirates
2
Production Control Department, Dubai Aluminum (DUBAL)
P.O. Box 3627, Dubai, United Arab Emirates
3
Technical Department, Dubai Cable Company (DUCAB)
P.O. Box 11529, Dubai, United Arab Emirates
(1 nbrahimi@sharjah.ac.ae, 2 abrar_khalaf@dubal.ae, 3 hebah.alhammadi@ducab.com)
*
important tasks and will give him/her flexibility to
Abstract - This paper presents an Operations Research incorporate urgent orders.
case study in an aluminum production company. The problem Production of aluminum is one of the fastest
under consideration consists of planning the production of industries in the world. Aluminum products are among the
billets obtained by cutting long aluminum logs produced in the most recycled materials after asphalt [1]. For an
cast house of the company. The objective is to minimize the
introduction to aluminum production see [2].
cost of recycling the scrap resulting from the casting and
cutting operations. The problem is a special case of the stock Billets are one form of aluminum products. They
cutting problem. Using integer linear programming modeling have a cylindrical shape and might have different alloys,
(ILP), the problem is solved for very large instances. Analysis diameters and lengths. The billets are cut from logs which
of the performance of the mathematical model was made are produced through a casting process by letting hot
through different scenarios and many data sets. The paper liquid aluminum flow through the molds of a casting
shows a real application of OR in industry and presents the table. Each table can produce a batch of several logs at the
first phase of a research project on this problem. same time.
The logs are cut using a saw which first eliminates
Keywords - Aluminum, Billet cutting, Cutting stock
the top crop of the log because it does not meet quality
problem, Integer linear programming, Metallurgy, scrap
requirements in terms of homogeneity. The length of the
I. INTRODUCTION crop is known for each diameter. After that, the billets are
cut based on three possible patterns: several short cuts of
The production of aluminum billets by cutting long the same length, one long cut, or a mix of one long cut
logs generates a lot of scrap with very high recycling with several short cuts of the same length. The
(mostly due to remelting) costs. When a cutting pattern is classification of billet cut lengths into short and long cuts
applied to an aluminum log, there is usually a remaining is made by the company. For confidentiality reasons, the
billet (scrap) whose length does not correspond to any values mentioned below do not necessarily correspond to
customer order. In order to minimize the total scrap, the the actual dimensions used in the company. Short billets
problem is formulated as a linear programming model have lengths between 200mm and 2000mm while long
taking into consideration the batch nature of billet billets are between 2010mm and 6000mm, both with
production. 10mm increment. The choice of these three patterns is
This paper presents the first phase of a research dictated by the characteristics of the saw. Before cutting,
project where more efficient methods are to be developed the casting length of the logs is between 3500mm and
to solve the problem over a longer period of time. 6200mm where all logs in one batch have the same
The authors have worked with Dubai Aluminum length. Table 1 summarizes the casting and cutting
(DUBAL), a big aluminum producer in the Gulf region, to lengths of the billets. After cutting the required billets, the
analyze the billet production process in order to develop last piece of the log is usually composed of a bottom crop
an efficient production planning tool that meets customer and a billet which does not correspond to any customer
demands while minimizing the total scrap generated in the order. During the cutting process, four types of scrap are
casting and cutting operations. Currently, this planning generated:
process is done manually which takes the planner several
days to finish one month plan. 1. Crops: Due to the starting and ending casting process.
Using an efficient integer linear programming (ILP) The length of crop depends on the log diameter (See
formulation and a MILP solver, good solutions were Fig. 1).
obtained in few minutes or few seconds for some 2. Cutting thickness: During the cutting operation, the
problems. This will allow the planner to focus on more saw blade deducts 10mm of the material, which will
be compressed and sent for recycling.
3. Extra billet: This corresponds to the extra piece of
*
Corresponding Author. billet which does not correspond to any customer
2,3
This work was done while the authors were students at order. It is usually recycled after a certain period of
the University of Sharjah time (Fig. 1).
4. Extra logs: The casting table is usually fully utilized difference between single stock-size and multiple stock-
and produces one full batch at a time. In some cases, size problems is that the former considers cutting from a
there are more logs produced than what is required. stock of the same size while the stocks can have different
They will undergo the same process as the scrap in sizes in the latter problem [9]. Using the four-field
the previous point. notation ( / / / ) introduced by Dyckhoff [8], the
closest notation to represent the problem is 1// / ,
TABLE 1 which means that we have a one dimensional cutting
LENGTHS AND CLASSIFICATION OF BILLETS stock problem (first field) with different figures for the
Description Range (mm) Increment (mm) assortment of large objects (D) and many items are to be
Casting lengths of logs 3500–6500 100 cut from many figures (M).
Cutting lengths Short 200–2000 10
In addition to the general characteristics of one-
of billets dimensional CSPs, the problem considered in this case
Long 2010–6000 10
study has some specific constraints. In particular, the
number of possible cut patterns can be classified into
The first and second types of scrap are unavoidable.
three categories: (i) one single long cut, (ii) Several short
However, a good plan should be able to minimize types
cuts (of the same length), (iii) Several short cuts (of the
three and four. The problem is a one-dimensional cutting
same length) with one long cut (See Fig. 1).Furthermore,
stock problem with special constraints and characteristics.
another type of scrap that is considered is composed of
In particular, the logs are produced in batches in the
the extra logs (stocks) that are produced in each batch.
casting operation. A mathematical programming
An integer linear programming formulation is presented
formulation was developed to minimize scrap.
and solved using a commercial solver. The results show
The remainder of the paper is as follows. In the next
that the problem with this size can be solved efficiently
section, the cutting stock problem is briefly reviewed and
using this approach. The authors are developing some
its relationship with our problem is presented. Section III
heuristic approaches to tackle larger problem instances.
presents the mathematical formulation. Characteristics of
the problem and the formulation are presented in Section
IV. Section V describes the experimental part and results.
The last section concludes this work and presents the next
steps in the research project.

II. MODELING AND SOLVING THE CUTTING

STOCK PROBLEMS

Gilmore and Gomory [3] define the cutting stock

problem (CSP) as the problem of filling an order at
minimum cost for specified number of lengths of material Fig. 1. The three possible groups of patterns: (i) one single
to be cut from given stock lengths in a given cost. One of long cut, (ii) Several short cuts, (iii) Several short cuts with one
the first researchers that highlighted CSP was long cut
Kantorovich in 1939 (The paper was published in 1960 as
III. PRESENTATION OF INTEGER LINEAR
[4]), when he discussed the best way of cutting a set of
PROGRAMMING FORMULATION
large objects into smaller items. Since then, many studies
on the subject of CSP were widely spread. A literature
Dimensions
review of the CSP can be found in [5] and [6].
Set of short billet cut lengths
McDiarmid [7] shows that the CSP is NP-hard in the
Set of long billet cut lengths
strong sense.
Set of lengths of logs from casting operation
The main objectives of solving the CSP are either to
Set of index for all logs for a given length
minimize the total cost of material or to maximize the
Data
utilization of the resources while satisfying demand and
Demand for billets with short cut length
technical constraints. Modeling and solving the CSP using
mathematical programming could be a good approach for Demand for billets with long cut length
solving some cases of the problem (See, for example, [3] cut length of short billet
and [10]). In the literature, both exact and heuristic cut length of long billet
approaches have been used to solve the CSP. Among the Thickness of the radial cutting saw (10mm)
exact methods we find dynamic programming (e.g. [11] Final length of logs produced with cast length
and [12]) and branch and bound procedures (e.g. [13]). in one cast.
Examples of heuristic methods include [14] and [15]. Batch size of logs wih length
Based on the typology of [9], the problem presented in Cost associated with one millemeter of scrap.
this paper can be closely identified as a one-dimensional Production cost of a batch of logs with length
multiple stock-size cutting stock problem, where the main
Decision variables: , 0,1 , , , , (13)
Quantity of short billet to be cut from log of
0 and integer, (14)
length
Quantity of long billet to be cut from log of 0,1 , , (15)
length
The objective (1) is to minimize the total cost of
Number of batches producing logs of length
recycling scrap due to cutting and to the production of
1 If 0
extra logs in each batch. The second part of the objective
0 Otherwise. leads to maximizing the utilization of the casting table. It
1 If 0 also allows the extension of the model to consider
0 Otherwise changeover cost for the table as logs of different
1 If log j of length l is used diameters require changing the table; which is a very
costly operation. This is presented in the research
0 Otherwise perspectives shown in the conclusion. Constraints (2)
The integer linear programming formulation is as
State that the total length of combined cuts should not
follows.
exceed the length of a log if the latter is used. Constraints
(3) and (4) link the integer variable and to the
binary variables and , respectively, by imposing
an upper bound on the number of billets (short or long)
that can be obtained from a log of a given length.
(1) Constraints (5), (6), and (7) indicate that there should be
at most one cut length of each category (short or long) on
a given log. Constraints (8) state that only one unit of a
long billet can be cut on one log. The demand satisfaction
of all billet lengths is guaranteed through Constraints (9)
and (10). Constraints (11) determine the number of
Subject to: produced batches. Finally, Constraints (12) to (15) are
non-negativity and integrity constraints.

IV. SOME PROPERTIES

, (2)
The following properties allow the classification of
the problem and provide the justification for the methods
that can be used to solve it.
/ , , , (3)
/ , , , (4) A. Problem complexity
This problem is an extension of the one-dimensional
2, , cutting stock problem (CSP). The one dimensional CSP is
(5)
NP-hard in the strong sense (See [7]). The proof of this
complexity is based on the reduction of the 3-
1, , PARTITION problem to the cutting stock problem. Thus,
(6)
it is better to develop heuristic solutions rather than exact
solution methods for large size instances of the problem.
1, , Most MILP solvers are all equipped with heuristics that
(7)
are able to quickly find feasible solutions for many hard
problems.
1, , , (8)
B. Dimensions of the mathematical programming
, (9) formulation
The ILP formulation presented above has the
following bounds (Table 2) on the different dimensions.
, (10)
TABLE 2
DIMENSIONS OF THE MATHEMATICAL
, (11) PROGRAMMING FORMULATION
Constraints 4
Integer variables
, 0 and integer, Binary variables
(12)
, , ,
 In order to reduce the size of the model, we made the The first feasible solution was obtained in less than
following changes. Firstly, the integer can be five seconds. However, the gap between this solution and
replaced with the binary variables since only one best lower bound was equal to 60.8%. After 20 seconds,
unit of a long billet can be cut on any log. This will the gap decreases below 40%. After that, it decreases
reduce the number of integer variables without increasing rapidly to 2.03% after 4 minutes. It is worth noting that
the number of binary variables. By eliminating the binary the decrease in the gap was mostly due to the
variables , constraints (8) must be deleted and improvement in the feasible solution. The improvement of
constraints (4) and (10) should be replaced with the lower bound value was very slow.
Constraints (4’) and (10’) below. B. Results with randomly generated data sets
In order to analyze the efficiency of the model, we
/ , , (4’)
have tested 120 data sets which were generated randomly.
The parameters that were considered are: batch size,
(10’) demand density, and problem size in terms of number of
possible lengths. Every time a parameter is varied
Secondly, the number of constraints can be changed, the other parameters are fixed at their default
considerably reduced in practice because among all values. For each combination of the three parameters, ten
possible billet length only few have positive demand. This instances were generated where the demands of each
is taken into consideration by adding the condition billet length is drawn from a uniform distribution between
0 and 0 to Constraints (3), (4’), (8), (9), et 10 and 100 units. Table 3 shows the values of varying
(10’). For example, Constraints (3) become: parameters and the default values of each parameter.
We have chosen three problem sizes: 1. Small, 2.
/ \ 0, , (3) Medium, and 3. Large. Demand density corresponds to
Finally, the variables associated with billet of zero the percentage of billet lengths whose demands are
demand are set to zero. positive. For example, a demand density of 0.15 means
that 85% of the billets will have no demand. Finally, the
V. NUMERICAL RESULTS batch size corresponds to the number of logs that can be
produced simultaneously during one casting operation.
The numerical results presented here are mostly based
on randomly generated data sets. For each instance, the TABLE 3
problem was solved using Xpress-MP version 2010. PARAMETERS OF THE DATA SETS
The mathematical programming formulation can be Parameter Value Default
solved using any MILP solver include open source solvers Problem size 1. 41 ; 50 ; 11 Size 2
such GLPK. The choice of Xpress-MP was mainly based 2. 81 ; 100 ; 21
on its availability and its capabilities. 3. 141 ; 150 ; 36
In the numerical results, the quality of obtained Demand density 0.15 ; 0.10 ; 0.05 ; 0.01 0.05
solutions is measured using the integrity gap between the Batch size 20 ; 30 ; 40 ; 60 ; 120 40
value of the heuristic feasible solution ( ) and the best
120, 100 , 500 for problems of size
lower bound ( ).
1, 100 , 1000 for size 2, and
% 100 200 , 2000 for size 3.
All numerical tests were carried on a personal 510 , 1000 for problems of size 1,
computer DELL OPTIPLEX 390 with Intel CORE i5 – 1010 , 2000 for size 2, and
3,10GHz microprocessor and 4Go of RAM. 2010 , 6000 for size 3.
1500 , 2500 for problems of size 1,
A. Results with the real data of the problem 3000 , 5000 for size 2, and
The model was initially solved using real data 4500 , 8000 for size 3.
collected from the cast house and planning department of 10 , 0.1, 100
the company. For confidentiality reasons we will not The maximum CPU time was set to 600 seconds.
present the data but will present the size of the problem in Table 4 shows the effect of problem size on the gap, the
addition to the CPU time and the quality of the obtained execution time and the number of obtained feasible
solution. solutions.
The problem consisted of planning the cutting of This table shows that even small size problems can be
billets of seven different lengths with the same diameter very hard to solve. The solver was not able to find the
and the same alloy. The billets can be obtained by cutting optimal solution of two small problems and the gap for
one or more of the 27 types of logs that the cast house can one of these two problems was very high (95.84%). We
produce. There are 310 lengths of short billets and lengths can conclude from the table that the bigger the problem
of long billets. the harder it is to find its optimal solution. The fact the
average gap does not depend a lot on the problem size
 might be explained by the choice of problem sizes and the and meta-heuristics to be applied on much larger
limited execution time. problems in the company. This approach will be mostly
useful when the problem is extended to multi-alloys,
TABLE 4 multi-diameters, and multi-periods.
EFFECT OF PROBLEM SIZE ON CPU TIME AND GAP.
Pb Avge Min Max Avge Min Max Opt.
Size Gap Gap CPU CPU CPU sols Acknowledgment
1 14.80% 0% 95.84% 139.2s 1.3s 600.5 s 8 The authors are very grateful to DUBAL personnel for
2 10.06% 0% 30.50% 377.2s 58.0s 602.2 s 5 their exceptional support with this project. We would like
3 12.31% 0% 29.74% 409.8s 140.5s 606.9 s 4 most particularly to thank Dr. Feras Allan, Vice President
of Product & Casting Operations- Cast house and Mr.
Demand density is a parameter related to the problem Mohammed Sweidan for their great help.
size since the number of constraints/variables depends on
the number of positive demands. Its effect on the quality REFERENCES
of the solution is presented in Table 5. For very small
demand densities, the problem becomes very easy to [1] Smith S., Asphalt Tops List of Most Recycled Materials,
solve. The solver finds optimal solution of the 10 EHS today, the magazine for environment, health and safety
instances in less than 250 seconds and in 42 seconds on leaders, 2003.http://ehstoday.com/news/ehs_imp_36326/
[2] Totten G.E., MacKenzie D.S., Handbook of Aluminum:
average. However, the results show that the average gap
Alloy Production and Materials Manufacturing, Edition
remain relatively small for large demand densities. We Marcel Dekker (2003).
believe that more tests need to be carried out to make a [3] Gilmore P.C., Gomory R.E., A linear programming approach
more clear idea of the relationship between demand to the cutting stock problem", Operations Research Volume
densities and the gaps. 9, PP 848-859 (1961).
[4] Kantorovich L., Mathematical methods of organising and
TABLE 5 planning production”, Management Science, Volume 6, PP
EFFECT OF DEMAND DENSITY 366-422, (1960).
Demand Avge Min Max Avge Min Max Opt [5] Haessler R.W., Sweeney P.E., Cutting stock problems and
density Gap Gap Gap CPU CPU CPU Sols solution procedures. European Journal of Operational
0.15 6.06% 0 24.33% 453.9s 42.1s 610.7s 5 Research, Volume 54, PP 141-150, (1991).
0.10 20.15% 0 50.72% 461.9s 68.2s 604.2s 3 [6] Ben Amor H., Valério de Carvalho J.M., Cutting Stock
0.05 12.64% 0 55.73% 346.5s 54.7s 602.9s 6 Problems in Column Generation, edited by Guy Desaulniers,
0.01 0 0 0 41.6s 1.1s 243.3s 10 Jacques Desrosiers, and Marius M. Solomon, Springer,
(2005).
[7] McDiarmid C., Pattern minimization in cutting stock
Finally, the effect of batch size on the solution is
problem, Discrete Applied Mathematics, Volume 98, PP
presented in Table 6. In general, the difficulty of the 121-130, (1999).
problem seems to increase with larger batch sizes. [8] Dyckhoff H., A typology of cutting and packing problems,
European Journal of Operational Research, Volume 44, No.
TABLE 6 2, PP 145-159 (1990).
EFFECT OF BATCH SIZE [9] Wäscher G., Haußner H., Schumann H., An Improved
Batch Avge Min Max Avge Min Max Opt Typology of Cutting and Packing Problems, European
size Gap Gap Gap CPU CPU CPU Sols Journal of Operational Research, Volume 183, No. 3, PP
20 4.61% 0 41.20% 212.8s 3.9s 603.6s 8 1109-1130, (2005).
30 25.96% 0 71.66% 439.6s 51.7s 603.9s 5 [10] Gilmore P.C., Gomory R.E., A linear programming
40 12.68% 0 55.73% 349.7s 55.2s 602.9s 6 approach to the cutting stock problem, Part II, Operations
60 26.78% 0 67.75% 564.3s 356.5s 603.5s 2 Research, Volume 11, 863-888 (1963).
120 31.97% 0 79.97% 410.3s 62.1s 603.9s 3 [11] Gilmore P.C., Gomory (R.E.), The theory and computation
of knapsack functions, Operations Research, Volume 14,
VI. CONCLUSION 1045-1074. (1966).
[12] Hifi M., Zissimopoulos V., A recursive exact algorithm for
weighted two-dimensional cutting, European Journal of
Integer linear programming (ILP) combined with a Operational Research, Volume 91, PP 553-564 (1996)
powerful solver can be a very good choice to solving real [13] Christofides N., Whitlock C. "An algorithm for two-
world industrial problems to which the aluminum billet dimensional cutting problems", Operational Research
cutting problem belongs. The planner in the company Volume 25, PP 30-44 (1977).
where this project was done used to spend several days [14] Berberler M.E., Nuriyev U.G. A new heuristic algorithm
working on the planning schedule, while the solver can for the one-dimensional cutting stock problem”, Application
of Computer and Mathematics, Volume 9, No. 1, PP 19-30,
find good feasible solutions in few minutes. (2010).
However, the use of MILP is usually not the best [15] Cui, Y., & Yang, Y. A heuristic for the one-dimensional
approach to solve the cutting stock problem and its cutting stock problem with usable leftover. European Journal
extensions since the CSP is NP-hard in the strong sense. of Operational Research, Volume 204, No. 2, PP 245–250,
Hence, we are working on simple construction heuristics (2010)