Control and Optimization ofSupply Chain Networks
B. Erik Ydstie Duncan Coffey Mark ReadCarnegie Mellon Dow Chemicals Elkem Metals
AIM : Formulate supply chain problem as a network (electrical circuit) and see what inferences we can draw.
GOAL: Self Optimizing Enterprise
“..supply chain is a network of organizations involved through upstream and downstream linkages in different products and processes.”
Christopher, 1998
Outline:
1. Motivating Example: Silicon production2.The Supply Chain as a Process Network3.The Self Optimizing Enterprise4.Decentralized SCM main results5.Centralized SCM (Perea-Grossman)6. Industrial Case study # 2: Glass manufacture 7.Conclusions
Hydro/FossilPower
Quartzite Coal Wood Logs Limestone/sandElectrodes
Pre-treatment/Screen/Weigh/Mix
Rx 1 Rx 2 … Rx N
Refining Casting ChemicalPrimary Al
Secondary Al
CrushScreen
Raw material and Power Suppliers
Grind
Chemical Aluminum and Construction Industry
Flui
d Be
d
Dis
tilla
tion
Me2SiCl2 +++
SiliconesMotor oils
Fillers
90%
Flui
d/Fi
xed
Bed
10%
MeClCu
HCl
Siemen’sProcessWafering
Oxidation
SiCl3
Si WafersElectronics
Photovolatics
Fumed SilicaZeolytesCatalysts
Silicon for Chemicals and Electronics
Melting/blending
Silicon for primary and secondary Aluminum
Aluminum ++
Ingot castingExtrusionRollingCasting
Al platesContainers
Transportation
SiO2+C+O2=Si+SiO2(a)+CO2
1. From Si Manufacture to Finished Products
MicrroSilica
Electronics $3 Aluminum $1-2 Silicone
Memory Chips $40 Airplanes $12 Motor-oilsProcessors $60 Cars (Audi 6,++) Silicone fillersPhotoVoltaics $20 Trains +++++ Beverage cans
Profiles $7(windows, siding)
1ppb 7% Six-Oy-Hz
Silicon Demand 1,016,000 Mt (2001)
42,000 MT Si-eq 401,000 MT Si-eq 573,000 MT Si-eq
Construction, transportation, consumer goodsExchange rate, production in China, EU regulations
Silicon by Market Sectors
• Clean, Renewable Energy from the SUN• Silicon: 14-16% Efficiency• 0.1 kWp per m2
• 100 kWp per ton Silicon• Government Incentives in
EU, Japan and US
Current world wide SG-Si market ~$30M
Financial Outlook: Silicon for PhotoVoltaics
On gridCentralized
On gridDecentralized
Off GridCentralized
Off gridDecentralized
Market Outlook Silicon for PhotoVoltaics
050
100150200250300350400450
1990
1993
1996
1997
1998
1999
2000
2001
2002
2003
MWpEG-Scrap
R&D Programs: ELKEMSGSBayer/DegussaDow CorningAstropowerSHARPScanWanfer++
12 % Growth
slowing
25-30 % Growth
accelerating
99%+ Silicon based
KPMG 1999: Economies of Scale and Better Production facilities can reduce Solar electricity generating cost by a factor of 3
The Supply Chain for EG-Si and SG-Si
Cryst Wafers Etch.TCS Dist CVDSiO2 Red. Silgrain
Metallurgical Si - Silgrain 99.5% Electronic Grade Si 1ppb 12,000 tons/year
Solar CellsWafersRemelt/Cryst
Solar Grade Si 1 ppm
Scrap Silicon - tops/tails/rejects 2000 tons/year
Elkem, … BayerMitsubushiASiMI,..
DegussaSharpAMDIntel,…
CrystalloxScanwaferSharp,…
AstropowerSharp, ShellBP, Kyocera….
DegussaKyoceraNEC,..
Missing link
$1 /kg $3 /kg $40-60/kg
$5-20 /kg
42,000 tons/year
2,500 tons/year
Value added = $2
2. Process Networks: A Brief Review.Program: Establish an iso-morphism between process systems and network theory.
Electricity
Coal
Quartzite
Electronics
1. topology2. transportation3. shipping/receiving4. manufacture5. storage6. forecasting7. performance evaluation
Background: Circuit AnalysisMeixner (Irreversible thermo.) 1960Perelson (Biology) 1970Alonso Coffey Ydstie (control, stability,..) 1997Hangos, Cameron, Perkins (modeling) 1999Gilles (modeling) 1998+++
All flow is caused by driving force (potential)
Aims: PDE’s replaced with ODE’sfocus on topologystability and controloptimizationdistribution of computationsease of modelingmodular software designvisualizationflows and ext. var. are additive++++
P
S S
S S S
S
Al(1) Al(2) Chemical
Heat and off gas Micro SilicaPowerQuartziteCoal
Electrical PowerElectrical Power
Quartz (SiO2), Carbon (C)Quartz (SiO2), Carbon (C)SiO, COSiO, CO
Silicon (Si(l))Silicon (Si(l))
(1+a) SiO2 + SiC → (1-a) Si(l) + (1-2a) SiO(g) + CO(g)(1+a) SiO2 + SiC → (1-a) Si(l) + (1-2a) SiO(g) + CO(g)
SiO(g) + 2 C → SiC + CO(g)
2a SiO → a SiO2 +a Si
SiO(g) + 2 C → SiC + CO(g)
2a SiO → a SiO2 +a Si
A A ASilicon Aluminum alloys
CastingRollingExtrusion,
Engine partsProfilesSheets++
Process ControlStabilization
Real Time Optimization
Supply Chain Management
S
P
TR
T
Assets A = {v,c,z}Activities (flow of assets) F = {f,z}
v – extensive variable (amount /charge)c – intensive variable (value/voltage)f – rate of flow of an extensive variable (flow rate/current)z – characterization (SKU, composition, etc.)
G=(P,S,T,R,H)
P – production/manufactureS – storageT – terminalsR – routingH – transportation
H
Activity
rate [units/time] , value generation [value/unit]
Σ fi=0
Material balanceKirchoff current law
S
dvdt = f
StorageCapacitor
InductorDiode+
Routing point Storage Activity and terminals
Rf=w
FourierFickNewtonOhm’s Laws
Constitutive equations
Conservation Laws Potential Flow
Activity Based Analysis – Value added as driving force
Each activity adds value (positive or negative)
Cos
t/Val
ue p
er u
nit $
Raw MaterialPurchase
Move and storeraw material
Produce Move and storefinished products
Sell and shipFinished products Activity
Purchase price
Added value w=cn-cn-1
Sale price
S
PH
Direction of flow: From low to higher value is positivefrom high to low value is negative
Net (internal) activity Cost:
Cost (internal) of operations:
Net rate of profit [$/sec]:
Circular activity does not add value: (Kirchoff’s voltage law)
Cost/dissipation: (2nd law of thermo)
Financial Implications:
-P plays the role of “energy”
G
f w
N-port
Diathermal
AdiabaticSem
i-per
mab
le T Sµ
N
-p
V
Thermodynamics Networks
Conservation laws Convex Energy function U(N,S.V)
Intensive variables via Legendre transform
Kirchoff’s Current LawKirchoff’s Voltage Law
U
f
Network builds itself into an energy minimizer!!!
G
c1 c2
Supplied energy:
Dissipation:
0 0.1 0.2 0.3 0.4-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
T-C
r=10w=1
f
w
One port example
G
c1 c2
Supplied energy:
Dissipation:
0 0.1 0.2 0.3 0.4-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
T-C
One port example
Generalizes to RLC and we get stability
3. The Self-Optimizing Enterprise
• One-ports (terminals, storage, production, activities, routers, +++)• Conservation of assets (KCL)• Circular activity does not add value (KVL)
f1 f2 f3 f4 f5 f6 f7
Distributed Enterprise
1. The enterprise consists of a (large) number of nodes.2. Connected with the “world” at terminal points (boundary conditions).3. Flows are the result of potential at the terminal.
Tellegen’s Theorem – a Topological Result
“Building the Self-Optimizing Enterprise”
• One-ports (terminals, storage, production, activities, routers, +++)• Conservation of assets (KCL)• Circular activity does not add value (KVL)
Distributed EnterpriseDistributed Enterprise (a)
Flows and costs are orthogonal
Enterprises (a) and (b) have same topology.Same enterprise in different states
Tellegen’s Theorem – a Topological Result
“Building the Self-Optimizing Enterprise”
• One-ports (terminals, storage, production, activities, routers, +++)• Conservation of assets (KCL)• Circular activity does not add value (KVL)
Distributed EnterpriseDistributed Enterprise (a)
Filtered lows and costs are orthogonal
Network operator – filter, forecast, Fourier transform, expectation,..
1. Existence and uniqueness of solutions if the routing policies are positive (negative).
2. Added value (cost) is stationary.3. Added value (cost) is maximized (minimized) if the routing
policies are negative.
Assumptions: Assets are conservedCircular activity does not add value
4. Decentralized SCM: Optimality Results
∆f
∆w ∆f
∆w
The activity cost is not positive
f
w
f
f
w
w(A) (B)
(C)
Capacity Constraint
f
w(D)
Discount for large volumes
Neutral costNegativemonotonicity
Positivemonotonicity
4. Decentralized SCM: Optimality Results
1. Existence and uniqueness of solutions if the routing policies are positive (negative).
2. Cost is stationary3. Passivity can be used to investigate stability..
Assumptions: Assets are conservedCircular activity does not add value
Load Balancing - Parallel Activities
1
2
wCooperative -stabilize Competitive - unstable
0 1 0 1
Competitive markets give narrow margins and pressure toreduce cost to stay in business.
Develop new products and new processes.Dominate market.Seek (cost) advantage (geographical, sourcing,…..
Cost Balancing - Serial Activities
1 2
w1
“Cooperative”
0 f
r1f1=w1 r2f2=w2
w2
w=w1+w2f=f1=f2
“Fair” transfer price::
w
Advantageous not to share information
(may be too limited)
Industrywide Cost Curve Silicon
Cos
t per
uni
t
World Wide Annual Supply 1BMt
Savings: $80M per year
• Better logistics and ERP systems• New energy and raw materials contracts• Better management of supply chain (smaller inventory)• Better scheduling and production planning• Better control and quality management (statistical methods, “6Sigma”)• Stream line business operation (lean manufacturing)
mean
Current market price
Motivating Case Study: Silicon Production
O. Sorli, CTO Elkem, 2000
1994
2000
Batch HouseBatch Mix
Hot End Cold End Cutting/Packaging
I/O I/O I/O I/OI/O
PC/Workstation PC/Workstation PC/Workstation
PLC -LogicStochastic ControlKalman Filter
PLC -LogicStochastic ControlEHACPID/FFMachine VisionNonlinear Adaptive Combustion AirNonlinear Compensation for Air Pressure Control
Nonlinear EstimatorTweel control Machine Vision
Optimal CuttingMachine VisionPLC-Logic
Technical IT AlgorithmsCommunication
Activities InventoryFlows
InterfaceActuatorsSensors
Crown I
Melter II
Refiner IVWaist III Canal V
Industrial Case Study #2: Glass Production
Quality checkLaminatingPackagingDelivery(MB/BMW)
Replacement parts
Activity
Actuators/Sensors
VirtualActivity
Smart Measurements
Raw materialand energy inputs
Molten glassand energy outputs
Virtual inputs Virtual outputs
Real Process
Virtual Process
RealMeasurements
VirtualMeasurements
Sensor Module
Distributed Systems thatIntegrate Physics and Communication
Distributed system of process activities
Sensor, actuator control
Network of computers, databases and software for control and optimization (Technical IT)
raw materials Intermediates and products
Accounting and finance (Business IT) Man/Machine Interface
Sensor, actuator control Sensor, actuator control
Control moduleOptimization module6Sigma module*****
Activities
Interface
Software
Estimated savings $35M per year
PW1
C1
A/B/C
D/E/F
G/H/I
A/B/C
D/G/H/I
A/D/F/G/I
D/E/F
C/G/H/I
C/F/I
C/E/G
G/H/I
B/E/G
A/F/G
A/B/C
D/E/F
D/E/F
G/H/I
D/E/F/G/H/I
A/B/C/D/E
F/H/I
A/D/E
B/C/E
A/D/G
Plants & Warehouses
Distribution Centers
Retailers Customers
A/B/C/D/E/F/G/H/I
A/B/C/D/E/F
A/F/G
C/E/F/G/I
C/D/E/F/G/H/I
A/D/F/G/H/I
A/B/C/D/E/F/H/I
A/D/G
A/B/C/D/E
B/C/E/F/G/H/I
A/B/C/D/E/F/G/H/ID/E/F/G/H/I
D/E/F
A/C/D/E/F/G/H/I
PW3
PW2
DC1
`DC2
DC3
DC4
RT1
RT2
RT3
RT4
RT5
RT6
RT7
RT8
RT9
RT10
C2
C4
C3
C5
C6
C8
C10
C12
C13
C14
C15
C18
C19
C20
C7
C17
C9
C11
C16
B/E/G/H/I
PW1
C1
A/B/C
D/E/F
G/H/I
A/B/C
D/G/H/I
A/D/F/G/I
D/E/F
C/G/H/I
C/F/I
C/E/G
G/H/I
B/E/G
A/F/G
A/B/C
D/E/F
D/E/F
G/H/I
D/E/F/G/H/I
A/B/C/D/E
F/H/I
A/D/E
B/C/E
A/D/G
Plants & Warehouses
Distribution Centers
Retailers Customers
A/B/C/D/E/F/G/H/I
A/B/C/D/E/F
A/F/G
C/E/F/G/I
C/D/E/F/G/H/I
A/D/F/G/H/I
A/B/C/D/E/F/H/I
A/D/G
A/B/C/D/E
B/C/E/F/G/H/I
A/B/C/D/E/F/G/H/ID/E/F/G/H/I
D/E/F
A/C/D/E/F/G/H/I
PW3
PW2
DC1
`DC2
DC3
DC4
RT1
RT2
RT3
RT4
RT5
RT6
RT7
RT8
RT9
RT10
C2
C4
C3
C5
C6
C8
C10
C12
C13
C14
C15
C18
C19
C20
C7
C17
C9
C11
C16
B/E/G/H/I
Material Flow(A, B or C)Information Flow(Orders)
Distr. Center
Plant Warehouse
Retailer CustomerPlant
ukk’’
t
ykc
dkc
Ik’,Ok’k ykk’’yk’k
uk’k
upk’Raw materials
supplier
yk`k``
uk`k``
ukk’’, upk’: Inputsdkc: DisturbancesIk,Okk’’: Statesykk’’,yk’’c: Outputs
Ik,Okk’’ Ik’’,Okc
Material Flow(A, B or C)Information Flow(Orders)
Material Flow(A, B or C)Information Flow(Orders)
Distr. Center
Plant Warehouse
Retailer CustomerPlant
ukk’’
t
ykc
dkc
Ik’,Ok’k ykk’’yk’k
uk’k
upk’Raw materials
supplier
yk`k``
uk`k``
ukk’’, upk’: Inputsdkc: DisturbancesIk,Okk’’: Statesykk’’,yk’’c: Outputs
Ik,Okk’’ Ik’’,Okc
5. Centralized SCM using MPC(Perea-Grossmann-Ydstie, 2002)
Plant and order delaysDiscrete manufactureMaximize profit
Solve using MPC strategy
Case 1: Optimal Integrated
0
5000
10000
15000
20000
1 13 25 37 49 61 73Time
Inve
ntor
y at
DC
1
D1I D2A D2C D2D D2E D2F D2G D2H
Case 3: Optimal Scheduling
0
5000
10000
15000
20000
1 13 25 37 49 61 73Time
Inve
ntor
y at
DC
1
D2A D2C D2D D2E D2F D2G D2H D2I
Case 1: Optimal Integrated
0
5000
10000
15000
20000
1 13 25 37 49 61 73Time
Inve
ntor
y at
DC
1
D1I D2A D2C D2D D2E D2F D2G D2H
Case 3: Optimal Scheduling
0
5000
10000
15000
20000
1 13 25 37 49 61 73Time
Inve
ntor
y at
DC
1
D2A D2C D2D D2E D2F D2G D2H D2I
• Better balance of plant schedule and inventory levels in CSM. • Centralized planning needed when the delays are long.• Too short planning horizon T < 6 days gives myopic policy (plant shuts down).• Policy insensitve to planning horizon for T .> 12 days.
Edgar Perea-Lopez, Grossmann Ydstie 2002.
GAMS / XPRESS-MP to solve MILP. Discrete time elements = 2hr. Prediction horizon of 12 days (144 elements of 2 hr each). Weekly updates updates of the demand, MILP model with 1,296 binary, 85,898 continuous variables and 59,150 constraints.
Summary• Practical control systems are built up from the bottom using
distributed modules.• Decentralized control and decision making can be “optimal” (self
optimization).• Parallel distributed computing is a reality. Numerical methods and
software is needed to take advantage.• Cutting cost vs new product and process.• Margins converge and become smaller over time in a competitive
commodity market.• Large, “hybrid” dynamic system solved to optimality (mpc).• Large savings possible• But, there is no “silver bullet”. Broad spectrum of technologies
need to be integrated with the process and Business IT system.
Acknowledgements: B. Marshall ALCOA Technical Center PghO. Sorli ELKEM ASAT. Halmo ELKEM ASA/StatoilY. Jiao PPG Inc Pgh