GLOBAL

OPTIMIZATION
Theory
Algorithms
Software
Applications
Nick Sahinidis
University of Illinois at
Urbana-Champaign
Department of
Chemical Engineering

OPTIMIZATION
in 30 seconds
f

x
optimum

Derivative must be zero at optimum

If constraints are present, use more
complex optimality criteria (KarushKuhn-Tucker conditions)

HISTORICAL
DEVELOPMENTS in
OPTIMIZATION
300 B.C: Shortest distance from a
point to a line (Euclid)
1600s: Leibniz/Newton (Calculus)
1847: Gradient methods (Cauchy)
1875: Minimum free energy
principle (Gibbs)
Late 40s: Linear Programming
Dantzig, ...
Linear objective function and constraints

Late 50s: Nonlinear Programming

Nonlinear functions
Wolfe, Abadie, Rosen, Fletcher, Davidon,
Reeves, Powell, Broyden, Fiacco,
McCormick, Rockafellar, Han, ...

60s: Integer Programming

Integer variables to model discrete
decisions and economies of scale
Beale, Benders, Gomory, Tomlin, Balas,
Geoffrion, Glover, ...

TODAYs
CHALLENGES
Stochastic Optimization
Hedging against uncertainty

Global Optimization
Problems with multiple-local minima

Computational Complexity
Approximation Schemes
Fast (Polynomial Time)
Near-Optimality Proofs

Applications
Supply Chain Management
Process and Product Design
Molecular Design and Analysis
Computational Chemistry
Bioinformatics
Software
General Purpose
Robust; Easy to use

THE MULTIPLE-MINIMA
DIFFICULTY
f

Classical optimality conditions are

necessary but not sufficient
Classical optimization provides the
local minimum closest to the
starting point used

PROCESS
SUPERSTRUCTURE

A network of processes and

chemicals.

Which processes to build or

expand in each time period?
Discrete decision

How much new capacity to

bring on-line?
Continuous decision

Mixed-integer (non)linear
programming formulations
(MINLP)

HEAT EXCHANGER
NETWORK SYNTHESIS
H1 @ 300

f3
t1
f1

f7

t2

H1 @ 320

f5

C1 @ 100
F = 10

C1 @ 280

f6

H2 @ 300
f2
f4
t3

t4

f8

H2 @ 340

Energy balances introduce

nonconvexities (conservation of
flow * temperature)

MOLECULAR
STRUCTURE PREDICTION
E

r ro
kb
ij
ij
ij
(ij)B

o
kijk ijk ijk

(ijk)

kijkl
(ijkl)

Aij
12
(ij)NB rij

k cos nijkl

ijkl

Bij

6
r
ij
r

qiq j
4 r
o ij
(ij)NB

Molecular Mechanics: Atomic

positions must minimize energy
difference from ideal state

MEDICAL DIAGNOSIS
& PROGNOSIS
Linear Separation of Sets:
A
B

Bilinear Separation of Sets:

A
B

A
B

A
A

B
A

UTILITY-BASED
PORTFOLIO OPTIMIZATION
UTILITY
RISK-AVERSE
RISK-NEUTRAL
RISK-SEEKING

WEALTH
UTILITY

GOAL

WEALTH

Problem Formulation
(P)

min f (x, y)

Objective Function

s.t. g(x, y) 0

Constraints

x Zp

Integrality Restrictions

y Rn

Continuous Variables

Challenges:
Multimodal Objective

111111111111111
000000000000000
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111
000000000000000
111111111111111

Objective

Integrality
Discrete Set

Nonconvex Constraints

Objective

Nonconvexity in
Feasible Space

MOTIVATION FOR
AN ALGORITHM
Diverse Application Areas

Combinatorial Chemistry
Molecular Biology
Finance
Parameter Estimation
Design of Experiments
Engineering Design
Process Synthesis
CAD
Layout Design

Theoretical Challenges
Important Subclasses
Mixed-Integer Linear Programming
Continuous Nonlinear Programming
Nonlinear 0-1 Programming

NP-hard Problem
Murty and Kabadi (1987)

MINLP ALGORITHMS
Branch and Bound
Bound problem over successively refined
partitions
Falk and Soland, 1969
Gupta and Ravindran, 1985

Convexification
Outer-approximate with increasingly tighter
convex programs

Tuy, 1964
Sherali and Adams, 1994

Decomposition
Project out some variables by solving
subproblem
Duran and Grossmann, 1986
Visweswaran and Floudas, 1990

Our approach
Branch and Bound
Convexification
Range reduction
Constraint Propagation
Duality-based

Branch and Bound Algorithm

Objective

Objective
P

U
R

L
Variable
a. Lower Bounding

Variable
b. Upper Bounding

Objective
P
R

U
L

R2
R1

R1

R2

Fathom
Variable
c. Domain Subdivision

Subdivide
d. Search Tree

FINITE vs. CONVERGENT

Branch & Bound Algorithms

Finite sequences

An infinite sequence
branch and bound methods applied to global
optimization problems generate infinite processes,
In contrast, in integer programming these are finite
procedures.
Horst & Tuy (1996)

Factorable Functions
(McCormick, 1976)
Denition: The factorable functions are recursive
compositions of sums and products of functions of
single variables.

Example: f (x, y, z, w) =

exp(xy + z ln w)z 3
x1 = xy

x2 = ln(w)

x5
x3

exp( xy + z ln w) z 3
x2

x1
x4
x7

x3 = zx2

x6
0.5

x4 = x1 + x3
x5 = exp(x4 )
x6 = z 3
x7 = x5 x6

f = x7

COMPUTATIONAL RESULTS
WITH STANDARD
BRANCH-AND-BOUND
Ex.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
*:

Ntot

Nopt

Nmem

3 1
2
0.8
1007 1
200 210
2122* 1 113* 1245*
17 1
5
6.7
1000* 1 1000* 417*
1 1
1
0.3 Ntot Total number of nodes
Nopt Node where optimum found
205 1
37 43
Nmem Max. no. nodes in memory
43 1
8
1
2192* 1 1000* 330* T
CPU sec (SPARC 2)
1 1
1
0.4
81 1
24 19
3 1
2
0.6
7 2
3
1.3
7 3
3
3.4
15 8
5
3.4
2323* 1 348* 1211*
1000* 1 1001* 166*
1 1
1
0.5
85 1
14 11.4
3162* 1 1001* 778*
7 1
4
1.2
9 1
4
1.2
75 6
13 11.7
7 3
2
1.5
17 9
9
2.9
Did not converge within limits of
T 1200 (=20 min), and N mem 1000 nodes.

CHALLENGES
Standard Global Optimization
Converges Slowly
Algorithmic Issues
Build tight relaxations
Reduce domain by automatically
identifying sub-optimal regions
Develop finite instead of merely
convergent algorithms

Implementation Issues
Existing NLP solves are often unreliable
Numerical instabilities are common with
nonlinear terms
Ease of use
Capability to sole problems of industrialscale magnitude

RANGE REDUCTION
Relaxed Value Function

x
xL

xU

If a variable goes to its upper

bound at the relaxed problem
solution, this variables lower
bound can be improved

PROBING
Q. What if a variable does not go to
a bound?
A. Use probing: temporarily fix
variable at a bound.

Relaxed Value Function

zL

zU
U
x

x*

xU

IMPLIED RESULTS
Monotone Complementarity Bounds
for Convex Programs
Mangasarian and McLinden
Mathematical Programming, 1985

Linearity-Based Tigthening
Andersen and Andersen
Mathematical Programming, 1995

Marginals-Based Range Reduction

Ryoo and Sahinidis
Journal of Global Optimization, 1996

Branch and Contract

Zamora and Grossmann
Journal of Global Optimization, 1999

SEPARABLE
CONCAVE MINIMIZATION

min f (x) = f k (x k )
k

f k (xk ) concave, k.
Ax b
L

x xx

LOWER BOUNDING
f(x)
g(x)
xL

[xL,xU]

xU

CONVEX ENVELOPE IS LINEAR

FINITE BRANCHING RULE

Variable selection:
Typically, select variable with largest
underestimating gap
Occasionally, select variable corresponding to
largest edge

Locating partition:
Typically, at the midpoint (exhaustiveness)
When possible, at the best currently known
solution
Makes underestimators exact at the
candidate solutions

Finite Isolation of global optimum

x2

x*
x1L

x1U

x1
d

xA

PROOF OF FINITENESS
g(x)

f(x)

Ascent directions of the concave

program (f) are also ascent directions
for the relaxation (g).
Once global optimum found, all open
nodes are finitely deleted.

TIGHT RELAXATIONS
f (x )
Concave over-estimator
Convex under-estimator

x
f (x )
Concave envelope
Convex envelope

x
f (x )

x
Convex/concave envelopes
often finitely generated

Convex Extensions of l.s.c. functions

Denition: A function f (x) is a convex extension of g(x) : C R
restricted to X C if
f (x) is convex on conv (X),
f (x) = g(x) for all x X.

Example: The Univariate Case

l

g(x)
m

f (x) is a convex extension of g(x)

restricted to {l, n, o, q}

f(x)
n

Convex extension of g(x) restricted to

{l, m, n, o, p, q} cannot be constructed

p
o

(0,0)

A Simple Univariate Example

3x

(2, 4)

Consider g(x) = x2 restricted to {0, 1, 2}.

Then,

x2

f (x) =
x

(0, 0)

(1, 1)
x

x
3x 2

x [0, 1]
x [1, 2]

as well as x2 itself form convex extensions.

The convex extension may in fact be tighter than the function

Constructibility of Convex Extensions

Theorem: A convex extension of g(x) : X R over a convex set
C X may be constructed if and only if
g(x) min

i g(xi ) |
i

i = 1, i (0, 1), xi X

i xi = x;
i

for all x X

where the above summations consist of nite terms.

An interesting sucient condition

If FI is an arbitrary collection of faces of a convex set C,

and we are interested in the convex extension of g(x)
restricted to FI over C,
Convex Extension
constructible over FI

Equivalence

g(x) is convex over

each face in FI

The Generating Set of a Function

Definition: The generating set of a function g(x) over a compact convex
set C is defined as
GC (g) =

x (x, y) vert epi conv g(x)

where vert() is the set of extreme points of ().

Examples:
g(x) = x2

g(x) = xy

x
x2
xy

G[0,6] (g) = {0} {6}

Convex Envelope

G[1,4]2 (g) = {1, 1} {1, 4} {4, 1} {4, 4}

Similar in vein to that for polyhedral convex functions in Rikun (1997)

Identifying the Generating Set

Characterization: x0 Gepi
C (g) if and only if there
exists X C and x0 Gepi
X (g).
Example I: X is linear joining (xL , y 0 ) and (xU , y 0 )

yU

x/y

Gepi (x/y) = (x, y) x {xL , xU }

xL
xU

yL

neighborhood of (x0 , y 0 )

Example II: X is

yU

x2 y 2
xL

xU

yL

Gepi (x2 y 2 ) = (x, y) x {xL , xU }

(x, y) y {y L , y U }

Identifying the Generating Set

Characterization I: x0 GC (g) if and only if there exists X C
and x0 GX (g).
Interesting Cases:
X is the relative interior of a line segment in C (Rikun 1997)
X = N (x), a neighborhood of x
C = C1 C2 and X is a slice isomorphic to C1 through x
Characterization II: The minimal subset of points that guarantee that
the optimal solution to the following problem:
g (s) = sup

s, x g(x) ,

remains the same for all values of s.

Example: Application to multilinear functions proves the polyhedral
nature of their extensions

Ratio: The Factorable Relaxation

z x/y
yU

x/y

xL

xL x xU

xU

yL

yL y yU

z x/y

cross-multiplying
U

yL y yU

xL /y U z xU /y L

x xx
L

y yy
L

zy x

xL x xU

Relaxing

z (xy U yxL + xL y U )/y U

z (xy L yxU + xU y L )/y L

Simplifying

zy (z xL /y U )(y y U ) x
zy (z xU /y L )(y y L ) x

yL y yU

yL y yU

xL x xU

xL x xU

Ratio: Convex Extensions

z x/y

0.5
2

2x4

2y4

x
4 2
Ratio: x/y

Convex Extensions yield Convex Envelope

yp zp xL
0.03
0
2

xU x

xU xL

yp (z zp ) y(z zp ) x

x xL

4 2

yp max

yp min

yU

x/y - Envelope

z zp , zp 0

xU x
xU xL
xU x
xU

xL

,y y

xU xL
L
U xx

, y yL

xU xL
x xL
xU xL

Factorable Relaxation is Weak

z 0.5 + x/4 + y/8

z 2 + x/2 y
4
0

y
x
4

x/y - Factorable

2x4
2y4

Ratio: The Concave Envelope

11
00
yU

x/y

Ghypo (x/y) = {xL , y L } {xL , y U }

{xU , y L } {xU , y U }

xL
xU

yL

Fit a Bilinear Function:

xL

1
(xU

xL )(y U

yL)

yL

(x x)(y y) +
U

xL
yU

(xU x)(y y L ) +

xU
xU
L
U
(x x )(y y) + U (x xL )(y y L )
L
y
y

Identical Concave Envelopes: (same generating set)

Ghypo (xy + xy L + xy U )/(y L y U ) =

y
x

{xL , y L } {xL , y U }
{xU , y L } {xU , y U }

xy + xyL + xyU
yL yU

Concave Envelope: (using bilinear envelopes)

1
yL yU

min y U x xL y + xL y L , y L x xU y + xU y U

Outer Approximation
Motivation:
Convex Programming Solvers are not robust
Linear programs can be solved eciently
Outer-Approximation:
Convex Functions are underestimated by tangent lines

(x)

xL

xU

Quadratic Convergence
with Projective Error Rule
Theorem:
Let (xj ) be a convex function over [xlj , xuj ] and
p the desired projective approximation error
Outer-approximate (xj ) at the end-points
At every iteration of the Sandwich Algorithm
construct an underestimator at the point that
maximizes the projective error of function with
current outer-approximation.
l
Let k = (xuj xlj )(xu
j xj )/

Then, the algorithm needs at most

0
k4
N (k) =

k2 ,
k>4
iterations

Branch And Reduce

Optimization Navigator
BARON CORE
Preprocessor
Data Organizer

USER MODULE

Range Reduction

Relaxation

Solver Links

Local Search

I/O Handler

Range Reduction

Debugging Facilities

Heuristics

Sparse Matrix Utilities

Heuristics

Core Module: Global Optimization Library

expandable, application-independent
Application Modules: Special Problem Classes
factorable nonlinear, linear multiplicative, bilinear,
concave, xed-charge, integer, fractional
programming
solve convex relaxations using CPLEX, OSL,
MINOS, SNOPT, SDPA

The Factorable NLP Module

Formulation:
min

f (x, y)
a gi (x, y) b

i = 1, . . . , m

U
xL

x
j
j
j

j = 1, . . . , n

yjL yj yjU

j = 1, . . . , p

xj real
yj integer
Form of fi (x, y) and gi (x, y):
f and g are recursive sums, products and ratios of
the following terms:
exponential, exp()
logarithmic, log()
monomial, ()a
Example: (Factorable Constraint)
1

x2 y 0.3 z
p2

+ exp

x2 p
y

xy 100

BARON ACCESS

On the internet since March 1995

http://archimedes.scs.uiuc.edu
Accessed from over 44 countries
10 accesses per day
ACCESS STATISTICS

US EDU
23%
Others
34%

US COM
17%

Europe
26%

Web interface since October 1999

Password-accessible online runs
Used by 135 research groups
Over 6000 problems solved

COMPUTATIONAL RESULTS
BRANCH-AND-BOUND

BRANCH-AND-REDUCE
No probing

Ex.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
*:

Ntot

Nopt

Nmem

With Probing

Ntot Nopt Nmem T Ntot Nopt Nmem T

3 1
2
0.8 1 1 1 0.5 1
1007 1
200 210
1 1 1 0.2 1
2122* 1 113* 1245* 31 1 7 20
9
17 1
5
6.7 3 1 2 0.4 1
1000* 1 1000* 417*
5 1 3 1.5 5
1 1
1
0.3 1 1 1 0.3 1
205 1
37 43
25 1 8 5.4 7
43 1
8 10
1 1 1 0.8 1
2192* 1 1000* 330* 19 1 8 5.4 13
1 1
1
0.4 1 1 1 0.4 1
81 1
24 19
3 1 2 0.6 1
3 1
2
0.6 1 1 1 0.2 1
7 2
3
1.3 3 1 2 0.7 1
7 3
3
3.4 7 3 3 2.7 3
15 8
5
3.4 1 1 1 0.3 1
2323* 1 348* 1211*
1 1 1 2.2 1
1000* 1 1001* 166*
1 1 1 3.7 1
1 1
1
0.5 1 1 1 0.5 1
85 1
14 11.4 9 1 4 1.8 1
3162* 1 1001* 778* 47 1 12 16.7 23
7 1
4
1.2 1 1 1 0.5 1
9 1
4
1.2 3 1 2 0.4 3
75 6
13 11.7 47 1 9 6.5 7
7 3
2
1.5 3 1 2 0.5 3
17 9
9
2.9 5 1 3 0.8 5
Did not converge within limits of
T 1200 (=20 min), and N mem 1000 nodes.
COMPUTER: SUN Sparc Station 2

1
1
1
1
1
1
1
1
1
1
1
1
1
3
1
1
1
1
1
1
1
1
1
1
1

1 0.7
1 0.3
5 48
1 0.3
3 2.4
1 0.3
2 5.8
1 0.8
4 7
1 0.4
1 0.7
1 0.2
1 0.7
2 3
1 0.3
1 2.4
1 4
1 0.6
1 1.4
5 15.4
1 0.5
2 0.5
4 5
2 0.6
3 1

CPU SEC FOR THE

PHILLIPS AND ROSEN
CONCAVE MINIMIZATION PROBLEMS

m
20
20
20
20
20
20
20
20
20
20
40
40
40
40
40
50
50
50
50
50

n
25
25
25
25
25
50
50
50
50
50
25
25
25
25
25
100
100
100
100
100

k
0
50
100
200
400
0
50
100
200
400
0
50
100
200
400
0
50
100
200
400

GOP (1993)
E=1%
(relative)

P&R (1990)
E=0.1%
(relative)

HP 730

CRAY 2
(parallel)
min
avg max
1
2
4
1
1
1
1
2
3
2
7
17
7
14
32
3
6
13
1
2
3
2
5
14
4
9
28
20
32
45

avg
0.5
2
17
33
82
0.6
17
47
109

std dev
0.01
2
20
28
58
0.01
31
49
80

0.5
1
3
25

0.02
0.6
4
26

BARON (1996)

=.000001
(absolute)
IBM RS/6000
Power PC
min avg
max
0.3
0.4
0.5
1
1
1
1
2
3
2
4
6
4
10
16
1
1
1
2
2.5
3
2
4
7
4
8
19
11
20
48
0.3
0.4
0.4
1
1
1
1
2
3
2
4
5
6
15
22
6
7
14
8
12
18
9
17
27
14
65
160
131
345
663

(m, n/k) = number of constraints, concave/linear variables.

HP 730 is 3-4 times faster than IBM RS/6000 Power PC.
CRAY 2 is 10+ times faster than IBM RS/6000 Power PC.

COMPUTATIONAL
RESULTS for
WDBC DATA
Wisconsin Diagnostic Breast
Cancer data:
ftp://ftp.cs.wisc.edu/math-prog/cpodataset/machine-learn/cancer1
2 classes: benign & malignant
9 attributes: clump thickness, uniformity
of cell size, uniformity of cell shape,
marginal adhesion, single epithelial cell
size, bare nuclei, bland chromatin,
normal nucleoli, mitoses
353 instances: 188 benign, 165 malignant
T raining Set
Size
150
176
235
353

BLP Size
(m,n,t)
(300,152,66)
(352,220,100)
(470,276,128)
(706,396,165)

%-Error in
T raining Set
0
0
1.7
2.5 **

%-Error in
T esting Set*
4
3
2.5
2.5 **

* Testing set consists of all 353 points

** WDBC data set is bilinearly inseparable

The Pooling Problem

f11

f21

3% S

x11
Blend X

$6
1% S

x21

Pool

2.5% S
$9
X 100

$16
x12

f12

2% S
$10

x22

cost
min
s.t.

Blend Y

X-revenue

1.5% S
$15
Y 200

Y -revenue

6f11 + 16f21 + 10f12 9(x11 + x21 ) 15(x12 + x22 )

q=

3f11 + f21
x11 + x12

f11 + f21 = x11 + x12

f12 = x21 + x22

qx11 + 2x21
x11 + x21
qx12 + 2x22
x12 + x22

Sulphur Mass Balance

Mass balance

2.5
Quality Requirements
1.5

x11 + x21 100

x12 + x22 200

Demands

Pooling Problems in Literature

Algorithm

Foulds 92

Computer
Linpack

Ben-Tal 94

GOP 96

BARON 99

BARON 00

CDC 4340

HP9000/730

RS6000/43P

RS6000/43P

> 3.5

49

59.9

59.9

**

106

106

Tolerance
Problem

Ntot

Ttot

Ntot

0.7

Haverly 2
Haverly 3

Haverly 1

Ttot

Ntot

Ttot

Ntot

Ttot

Ntot

Ttot

12

0.22

0.09

0.09

12

0.21

0.09

0.13

14

0.26

0.13

0.07

Foulds 2

3.0

0.10

0.04

Foulds 3

10.5

2.33

1.70

Foulds 4

25

125.0

2.59

0.38

Foulds 5

125

163.6

0.86

0.10

Ben-Tal 4

25

0.95

0.11

0.13

Ben-Tal 5

283

41

5.80

1.12

1.22

Example 1

6174

425

15

12.30

Example 2

10743

1115

19

8.56

Example 3

79944

19314

13

18.30

Example 4

1980

182

5.12

* Blank indicates problem not reported or not solved

** 0.05% for Haverly 1, 2, 3, 0.05% for Ben-Tal 4 and 1% for Ben-Tal 5

Molecular Design Problems

(Example: Refrigerant Design)

CH3
Cl

CH2
N

Cl

CH2

CH

Cl

Cl

CH2

Combinatorial Choice
F
Freon (CCl2F2)
F

Cl
Cl
Property Prediction

Satises
Property
Requirements?

Automotive Refrigerant Design

(Joback and Stephanopoulos, 1984)
Condenser
Tcnd
Expansion
Valve

Tavg = (Tcnd + Tevp )/ 2

Compressor

Tevp
Evaporator

max
s.t.

Hve
Cpla
Hve 18.4
Cpla 32.2
Pvpe 1.4
Property Prediction Constraints
Structure Feasibility Constraints
Nonnegativity Constraints
Integrality Requirements

High enthalpy of vaporization Hve reduces the

amount of refrigerant
Lower liquid heat capacity Cpla reduces amount
of vapor generated in expansion valve

Functional Groups Considered

Acyclic Groups

Cyclic Groups

Halogen Groups

Oxygen Groups

Nitrogen Groups

Sulfur Groups

CH3

CH2 r

OH

NH2

SH

CH2

r
r

> CH r

Cl

> NH

> CH

> CH r

Br
I

O r

r
r

> NH

>C<

r
r

> C < rr

= CH2

> C < rr

r
r

> C < rr

CHO

= CH r

COOH

CN
NO2

= CH
=C<

> CO

> N

> CO

= N

=C=

= C < rr

COO

CH

= C <r

=O

S r

= N r

= C < rr

Number of Groups = 44
Maximum Selection Size = 15
Candidates = 39, 895, 566, 894, 524

Property Prediction Constraints

N

Tb = 198.2 +

ni Tbi

= 15.2518

i=1

Tb

Tc =

N
i=1

0.584 + 0.965

N
i=1

ni Tci (

ni Tci )2

Pc =

(0.113 + 0.0032

N
i=1

N
i=1

n i ai

ni Pci

1
4.1868

Cp0a + 8.314 1.45 +

i=1

17.11 + 25.2

ni Cp0bi + 0.21 Tavg

i=1

ni Cp0ci 3.91 10

2
Tavg

Hvb = 15.3 +

ni Cp0di + 2.06 10

3
Tavg

Tavgr
Tcndr
Tevpr

+ 0.25

1.742
1 Tavgr

ni Hvbi

Hve = Hvb

i=1

Tbr =

Tavgr

1 Tavgr

i=1

i=1
N

(1 Tavgr )1/3

0.45

13.4721ln(Tbr ) + 0.43577Tbr6

Cpla =
)2

Tbr

ni Cp0ai 37.93 +

Cp0a =

15.6875

Tb

h=

Tc
Tavg
=
Tc
Tcnd
=
Tc
Tevp
=
Tc

= 5.97214 ln
0.169347Tbr6

1 Tevp /Tc

0.38

1 Tb /Tc
Tbr ln(Pc /1.013)
1 Tbr

G = 0.4835 + 0.4605h
k=

Pc
1.013

6.09648
Tbr

+ 1.28862ln(Tbr )

h/G (1 + Tbr )

(3 + Tbr )(1 Tbr )2

G
2
ln Pvpcr =
+ k(3 + Tcndr )(1 Tcndr )3
1 Tcndr
Tcndr
G
2
ln Pvper =
+ k(3 + Tevpr )(1 Tevpr )3
1 Tevpr
Tevpr
ni integer

Structure Feasibility Constraints

N

ni 2

ni 1 if

i=1

iDSR

YA

ni Nmax YA A

ni = 2ZB ;

iA

YC

iB

ni Nmax YC C

iDR

YA + YC 1 YM YA + YC

ni = 2ZT
iT

ni Nmax YR R

ni (2 bi ) = 2m

iR
N

ni bi 2(
i=1
N

i
N

ni 1);
i=1
N

ni bi

ni nj (bj 1) + 2

ni

i=1
iO,

ni 1 if
iSD

ni 1 and
iS/D

ni 1 if
iST

iSSR

ni 1 if

single-bonded

iD/S

iS/T

ni 1 if

ni 1

ni 1 and

j = 1, . . . , N

i=1

ni 1

ni
i=1

iSDR

ni = 2ZSR
iSR

ni = 2ZDR

iM

i=1

ni = 2ZS ;
iS

iD

ni Nmax YM M

3YR

ni 1
iSR/D

ni = 2ZD

iC

YM

ni 1 and
iD/SR

ni Sai if
iH

ni

ni 1

iO,

double-bonded

ni = 0
iH

ni Dai if
iH

ni = 0
iH

iT /S

ni 1 and
iS/SR

ni 1 and
iS/DR

ni 1
iSR/S

ni
iO,

triple-bonded

iH

ni = 0
iH

ni (Sai + Dai + Tai )

ni 1
iDR/S

ni Tai if

iH

ni 2(
iO

ni 1)
iH

Molecular Structures
Molecular Structure

Hve
Cpla

ClNO

(Cl)(N =)(= O)

1.5971

IF

(I)(F)

O3

(r O r )3

Molecular Structure

Hve
Cpla

CH2 FN

(F)(N =)(= CH2 )

1.1177

1.5297

ClF

(Cl)(F)

1.0480

1.5219

CClFO

(O =)(= C <)(Cl)(F)

1.0179

1.4855
1.3464

FNO

(F)(N =)(= O)

0.9893
ClFO

(Cl)(O)(F)

1.2880

0.9822
0.9342

1.2767

C3 H4

(CH3 )(C )( CH)

0.9283

1.2594

C2 F2

( C)2 (F)2

0.9229

CHClO

(Cl)(CH =)(= O)

1.1804

C3 H6 O

(CH3 )2 (= C <)(= O)

0.8978

FSH

(F)(SH)

1.1697

C3 H3 FO

(O =)(= CH)3 (F)

0.8868

1.1653

CF2 NHO (O =)(= C <)(F)2 (> NH) 0.8763

1.1574

C2 H6

1.1219

CHFO3

C2 HClO2 (= C <)(= CH)(Cl)(= O)2 1.1207

C3 H3 F3

CH3 Cl

(CH3 )(Cl)

All 44 feasible molecules identied in <

1
2

hour

(CH3 )2

0.8632

(O =)(= CH)(O)2 (F) 0.8288

(r > CH r )3 (F)3

0.5977

For CCl2 F2 , Hve /Cpla 0.57

Reduce

Isolate

Convexify

x*

BRANCH-AND-REDUCE
Engineering
design

Supply chain
operations
Molecular
design