Cyberinfrastructure-enabled Molecular Products Design...
Transcript of Cyberinfrastructure-enabled Molecular Products Design...
1
Cyberinfrastructure-enabled Molecular Products Design
and Engineering: Challenges and Opportunities
Venkat VenkatasubramanianLaboratory for Intelligent Process Systems
Purdue UniversityWest Lafayette, IN, USA
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 2
LIPS
OutlineMolecular Products Design
Data, Information and Knowledge Modeling Challenges
Cyberinfrastructure Ontological Informatics
Industrial Case StudiesLubrizol: Fuel Additives DesignCaterpillar: Rubber Products DesignExxonMobil: Catalyst DesignEli Lilly: Seromycin Formulation for MDR-TB
Summary
Venkat Venkatasubramanian, Plenary Lecture, Foundations of Computer-Aided Process Design, Princeton, July 2004.
Acknowledgements
Prof. J. F. Caruthers and Prof. W. N. DelgassDr. K. Chan (Bear Stearns), Dr. A. Sundaram (ExxonMobil), Dr. P. Ghosh (ExxonMobil), Dr. S. Katare (Dow), Dr. A. Sirkar (Intel), Dr. P. Patkar (ZS), Dr. S-H. Hsu (Purdue), S. Syal (Purdue), B. K. Balachandra (Purdue)
Funding AgenciesNSFDOEIndiana 21st Century Research and Technology FundGE, Bangalore, IndiaLubrizolCaterpillar ExxonMobilEquistarCummins
Materials Design
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 4
LIPS
Acknowledgements
Dr. C. Zhao (Bayer)Dr. A. Jain (United)Dr. S-H. Hsu (Purdue)L. Hailemariam (Purdue)P. Akkisetty (Purdue)P. Sureshbabu (Purdue)I. Hamdan (Purdue)M. Chaffee (Purdue)A. Shukla (Purdue)
Funding Agencies:NSF ERCIndiana 21st Century R&T FundEli LillyPfizerAbbott
Dr. Prabir BasuDr. Girish JoglekarProf. Ken MorrisProf. G. V. ReklaitisInformation Technology@Purdue
Pharmaceutical Engineering
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
LUBRIZOL: Fuel Additive Design
EPA requirement: Minimize intake-valve deposits (IVD)Approach: Fuel Additives Performance measure
BMW Test for IVD Stipulated to be less than 100 mg over a 10,000 mile road test
Expensive and time-consuming testing
Around $8000 for a single datum
Problem: Design fuel-additives that meet desired IVD performance levels
Intake Valve and Manifold
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
About 1000 Rubber Parts in Failure Critical Functions
Tires, Treads, Hoses, Shock Absorbers, O-rings, Gaskets, Mounts …
Reliability and Warranty Problems
CATERPILLAR
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
QuantumChemistry
T̃ = 2ρR∂ ψ∂I 1
+ I 1∂ ψ∂I 2
⎧⎨⎩
⎫⎬⎭
I −∂ ψ∂I 2
C + I 3∂ ψ∂I 3
C−1⎡
⎣⎢⎤
⎦⎥••∂ C∂C
Constitutive Model
3D FiniteElementPart Model
Spring-DashpotApproximationwith ChemicalAging
Ax + S8k1⎯ →⎯⎯ Ax+8
Ax + R k2⎯ →⎯⎯ Bx
Bxk3⎯ →⎯⎯ Bx
*
etc.Time
To rq u eCure Curve
Kinetics Polymer Network Mechanical Response
Mixing+
Heat Transfer+
Mold Filling
Part Manufacturing
Overcure
Undercure
SC
N S SS
SS
S
S
S
Zn
NewMolecules
MaterialFormulation
Manufac-turing
Process
PartShapeDesign
SubAssemblyDesignFull
ProductDesign &Operation
Design &Operation ofComplexSub-Assembly
Engine
Over 1000 rubber parts infailure critical functions
Multi-Scale Modeling
Quantum Mechanics
Kinetics
Heat Transfer FEA
Mechanics FEA
Part Design CAD /CAM
Caterpillar’s Multi-Scale Modeling Challenge: Design of Formulated Rubber Parts
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
ExxonMobil’s Grand Challenge: Catalyst Design via Combinatorial Chemistry
Fundamental Chemistry
Reactionkinetics
Performancemeasures
time
Rat
e/S
elec
tivity
C
k 6
A
B
F
k1
HD
R
M
E G
N
k 3 k4k2
k7
k5
Reaction network
Electronic/Structural descriptors
d band centerElectronegativity
...
Quantum calculations
Catalyst StructuresSynthesis
Characterization
10/11/2005
9
Pharmaceutical Product Development and Engineering
• Two ProductsDrugsDocuments
Preliminary product recipe
Drug is converted into ParticlesDrug Synthesis
Raw Chemicals
Formulation
Process Development
Preliminary process recipe
Scale/up -Tech Transfer
Adjusted process recipe
Manufacturing
A Delayed Product
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Molecular Products DesignGiven a set of desired performance specifications, how does one rationally and efficiently identify the optimal product structures and formulations?$200+ Billion/yr industry: Major Opportunity for ChEsAreas of Application
Engineering MaterialsFuel and Oil AdditivesPolymer CompositesRubber CompoundsCatalystsSolvents, Paints and Varnishes…..
PharmaceuticalsDrug Design
Agricultural ChemicalsPesticidesInsecticides
Zeolite!!
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Different Products, But Common Challenges
Lots of DataUncertain/Noisy Data Complex chemistry, lots of speciesNonlinear systems and processesIncomplete, Uncertain, MechanismsCombinatorially large search spacesNeed Multi-scale ModelsHundreds of Differential and Algebraic Equations to formulate and solveLaborious and time consuming model development processLimited human expertise
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Traditional Approach
Revise Design
No
Designer
Hypothetical Molecule
Synthesize
Evaluate
Design Objectives
Yes
Solution
Meets Design Objectives ?
Drawback: Protracted and Expensive Design CycleNeed a Rational, Automated Approach: Discovery Informatics
EUREKA!!!
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 13
LIPS
Need a New ParadigmCyberinfrastructure Methods and Tools
Ontology-based Discovery Informatics
Venkat Venkatasubramanian, Plenary Lecture,Chemical Process Control 7, Canada, Jan 2006 14
LIPS
Data, Information, and KnowledgeData
What’s going on?Raw, does not inform much
InformationHow are the variables related?
Correlations and relationshipsKnowledge
Why are they related?Develop Mechanistic understandingFirst-Principles Based Math ModelsHeuristics
Tower of Babel of Data, Information and Models
Venkat Venkatasubramanian, Plenary Lecture, Pharmaceutical Sciences World Congress, Amsterdam, April 2007 15
LIPS
What is an Ontology?Originated in Philosophy
Study of ExistenceNot to be confused with Epistemology: Theory of Knowledge and Knowing
Computer Science and Artificial IntelligenceA formal, explicit specification of a shared conceptualizationKnowledge representation in AI
Logic, Semantic Networks, Frames, Objects, OntologyWeb-driven development
Semantic Richness: Description Logic (DL) provides foundation for formal reasoningEasily create, share and reuse knowledgeSemantic Search
Venkat Venkatasubramanian, Plenary Lecture, Pharmaceutical Sciences World Congress, Amsterdam, Apr 2007 16
LIPS
Based on Set Theory and First-order Logic
An ontology is a 5-tuple ),,,, oc ArelHRC(:=Ο consisting of • Two disjoint sets C and R whose elements are called concept identifiers and
relation identifiers, respectively • A concept hierarchy CH : CH is a directed, transitive relation CCH C ×⊆ which
is called concept hierarchy or taxonomy. ),( 21 CCH C means that 1C is a subconcept of 2C .
• A function CCRrel ×→: that relates concepts non-taxonomically. The function dom: CR→ with ))((:)( 1 RrelRdom Π= gives the domain of R , and range:
CR→ with ))((:)( 2 RrelRrange Π= give its range. For ),()( 21 CCRrel = we also write ),( 21 CCR .
• A set of axioms oA , expressed in an appropriate logic language, e.g. first order logic
17
Ontology: A Simple Example
Excipient Properties
Material
API
Is aIs a
has
Concept Instance Property
RolesComposition
Range
hashas
Flow aid
Filler
Lubricant
Minimum
Maximum
Average
Is an instance of has
Contact angle
Moisture content
Is aMaterial
Composition
has
Web Ontology Language: OWL
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Catalyst Design Problem – The Grand Challenge
Fundamental Chemistry
Reactionkinetics
Performancemeasures
time
Rat
e/S
elec
tivity
C
k 6
A
B
F
k1
HD
R
M
E G
N
k 3 k4k2
k7
k5
Reaction network
Electronic/Structural descriptors
d band centerElectronegativity
...
Quantum calculations
Catalyst StructuresSynthesis
Characterization
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
1 A° 100 A° 10 μm 1 cm10-15 s
10-12 s
10-6 s
1 s
102 s
Quantum Mechanics
DFT
Atomistic SimulationMolec. Dyn.
Statistical Mechanics
TST
Continuum Models
Transport Models
Elect.Struct. Thermo.
Props.
Molec. Struct.
Rate Consts.
Overall Rxn Rates
Conc.Profiles
Microscopic Mesoscopic Macroscopic
Modeling at different Length and Time Scales
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Paraffin Aromatization
Identify a catalyst formulation for light paraffin aromatization that is superior to Ga/H-ZSM-5 in terms of:
Higher Benzene, Toluene, Xylene(B/T/X) selectivityHigher Hydrogen selectivity
Microkinetic model development for the kinetic description of the system
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Previous Work
Microkinetic Analysis (Dumesic, 1993)First unified view of reaction engineering on catalytic surfacesIncomplete forward problemLacks clear view of the design perspective (inverse problem)
Empirical Catalyst Design (Baerns, 2000)First attempt to “design” catalystsCompletely based on “guided” experiments – time consuming and expensiveNo fundamental understanding of the system (forward problem)
Ertl, Rasmussen, Lauterbach: Surface Science StudiesSteve Jaffe: Composition based modeling of large systemsJens Norskov: Computation based studiesSymyx, Novodynamics, Lauterbach: Combinatorial HTE
Venkat Venkatasubramanian, Plenary Lecture, Foundations of Computer-Aided Process Design, Princeton, July 2004.
Reaction Network – Propane Aromatization on HZSM-5
• 31 gas phase species + 29 surface species + 271 reaction steps• Model with 31 ODEs, 29 algebraic equations • 13 parameters with up to 10 orders of magnitude bounds on each
Hydride transfer
Cx-y=β-scission/Oligomerization
Cx+ Cy
+
H2 Cx-y
Dehyd
rogen
ation Protolysis
Cx=
adso
rptio
n/de
sorp
tion adsorption/desorption
Cy=
H+
CxH+
Cx
initiation
Paraffin
Dehydro-cyclization
Aromatics
Cx ParaffinCx
= OlefinPx
= Poly-olefinCx
=c Cyclic-olefinActive siteH+
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Ontological Informatics: Modeling Super Highway
• DAE System: 30 to 1000 species and 10 to 50 parameters• Parameter search space is highly nonlinear with numerous local minimaGA based pseudo global parameter estimatorFast and efficient coverage of the parameter space
• Optimization: Traditional least-squares and features-based objectives
ChemistryCompiler
Reaction Description Language Plus (RDL+)
EquationGenerator
Automatic generation of differential algebraic
equations (DAEs)
Parameter Optimizer
Solution of DAEsLeast squares
Features
Advanced Statistical Analyzer
Sensitivity AnalysisUncertainty AnalysisError Propagation
ReactionsGrouping
Chemistry Rules
time
Rat
e/S
elec
tivity
Performance Curves
Data
J.M. Caruthers, J.A. Lauterbach, K.T. Thomson, V. Venkatasubramanian, C.M. Snively, A. Bhan, S. Katare and G. Oskarsdottir, “Catalyst Design: Knowledge Extraction from High Throughput Experimentation”, In “Understanding Catalysis from a Fundamental Perspective: Past, Present, and Future”, A. Bell, M. Che and W.N. Delgass, Eds., Invited Paper for the 40th Anniversary Issue of the Journal of Catalysis, 2003.
Katare, S., Caruthers, J. M., Delgass, W. N., Venkatasubramanian, V., “An Intelligent System for Reaction Kinetic Modeling and Catalyst Design”, Ind. Eng. Chem. Res. and Dev., 2004.
Model Refinement Procedure
Crude modelsInitial data
Model screening
Infer new knowledge about models (mechanism) and data
Model discriminationModel enhancement
Enhanced modelsRicher data set
Planning new experiments
Add/Delete models
Add/Delete data
Model analysisData analysis
Chemistry Rules for Propane Aromatization on HZSM-5
Chemistry Rules Representative Chemical Reactions
1. Alkane adsorption
2. Alkane desorption
3. Carbonium ion protolysis
4. Carbonium ion dehydrogenation
5. Olefin adsorption
6. Olefin desorption
7. Aromatization
8. Beta-Scission
9. Hydride Transfer
10. Oligomerization
CH4 +
+ H2
+
+ +
+
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Modeling Super Highway: Reaction Modeling Suite (RMS)
Reaction Network
Reaction Description Language PlusEnglish Language Rules
C
k 6
A
B
F
k1
HD
R
M
E G
N
k 3 k4k2
k7
k5
Mathematical Equations
1//
541
1
=++−+=
−=
CBA
AB
AA
BkDkCkdtdCCkdtdC
θθθ
100’s of DAE’s
...
Model Generator
Beta ScissionLabel-site c1+ (find positive carbon)Label-site c2 (find neutral-carbon attached-to c1+)Label-site c3 (find neutral-carbon attached-to c2)Forbid (primary c3)Forbid (less-than (size-of reactant) 9)Disconnect c2 c3)Increase-order-of (find bond connecting c1+ c2)Add-charge c3Subtract-charge c1+
Beta ScissionLabel-site c1+ (find positive carbon)Require (c1+ primary and product)set-k k1Label-site c2+ (find positive carbon)Require (c2+ secondary and product)set-k 20*k1Label-site c3+ (find positive carbon)Require (c2+ tertiary and product)set-k 60*k1
Chemistry
8. Beta Scissiontransforms a carbenium ion into a smaller carbenium ion and an olefin
Grouping
8. a. Formation of a secondary carbenium ion
is 20 times faster than a primary carbenium ion
b. Formation of a tertiary carbenium ion
is 60 times faster than a primary carbenium ion
...
...
...
...
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Results
Over prediction of C2s and under prediction of aromaticsSpace time x 104 [hr]
Wt %
C2H6 C3H6
C3H8 Aromatics
Data*
* Lukyanov et. al., Ind. Eng. Chem. Res. , 34, 516-523, 1995
Model
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Results – Model Refinement
Refined model adds alkylation step to convert lighter alkanes to higher onesSpace time x 104 (hr)
Wt %
C2H6 C3H6
C3H8 Aromatics
DataOriginal ModelRefined Model
GA Based Pseudo Global Parameter Estimator
S.no. Name Esposito & Floudas (2000) Global optimizer
GA based procedure
υ Objective CPU s reported
min max
υ Objective Time taken
(CPU s)
Scaled CPU s
Time taken *671/211 (% Saving)
1 First order
irreversible chain reaction
5.0035 1.0000 1.18584x10-6 2.92
20.05 5.0122 0.9976 9.25418x10-6 0.24 0.76 (74)
2a First order
reversible chain reaction
4.0000 2.0000 40.013 20.007
1.8897x10-7 568.44 6164.3
3.9813 1.9759 39.787 19.887
8.1313x10-6 0.45 1.43 (100)
2b Same as 2a but with error in data
4.021 2.052 39.45 19.62
1.586x10-3 272.91 1899.82
4.001 2.027 39.22 19.50
1.589x10-3 0.5 1.59 (99)
3 Catalytic
cracking of gas oil
12.214 7.9798 2.2216
2.65567x10-3 79.77 1185
12.246 7.9614 2.2351
2.68017x10-3 0.38 1.21 (98)
4 Bellman’s problem
4.5704x10-6
2.7845x10-4 22.03094 36.16 12222
4.5815x10-6 2.7899x10-4 22.2885 3.54 11.26 (69)
5 Methanol-to-hydrocarbon
process
5.1981 1.2112
0 0 0
0.10652
1361.8 19125
5.2212 1.2320
1.6363x10-31 4.0059x10-12
0.004665
0.10586 2.08 6.61 (100)
6 Lotka-Volterra Problem
3.2434 0.9209 1.24924x10-3 367.26
9689.67 3.1434 0.9583 2.39784x10-3 0.56 1.78 (100)
Our zeolite model31 ODEs, 29 Algebraic equations
13 parameters9 concentration curves
Performance comparison on test problems – worst case: 3 ODEs/5 parameters
Reasonable comparisonto experimental data
2 hours on a Sun 400 MHz solaris machine
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Ontological Informatics for Catalyst Design: Summary
Chemistry Rules
High Throughput Experimental
Data
Reaction Knowledge
Base
Kinetic Modeling Toolkit
(Reaction Modeling Suites)
Chemistry Rules
High Throughput Experimental
Data
Reaction Knowledge
Base
Kinetic Modeling Toolkit
(Reaction Modeling Suites)
Represents reaction chemistries explicitlyCustomizable for different catalyst chemistries Results
Real-time evaluation of 100s of reaction pathways and 1000s of DAEsSaved months of model development effortImproved mechanistic understanding and first principles models
Molecule Ontology
Reaction Mechanism
Ontology
Reaction Ontology Chemistry
Knowledge Base
Molecule Classification
Reaction Classification
Reaction Network
Generator
Correctness and Consistency Checking
Proposed Mechanism
Graphical User Interface
Chemistry Ontology
Application
Ontology
Molecule Ontology
Reaction Mechanism
Ontology
Reaction Ontology
Molecule Ontology
Reaction Mechanism
Ontology
Reaction Ontology Chemistry
Knowledge Base
Molecule Classification
Reaction Classification
Reaction Network
Generator
Correctness and Consistency Checking
Proposed Mechanism
Graphical User Interface
Chemistry Ontology
Application
Ontology
Application
Proposed Framework for Knowledge Extraction from HTE – Kinetic Modeling
Reaction Modeling Suite
Automatic processing of data
Reaction Modeling Suite
Automatic processing of data
Model RefinementRules Features mapping
Suggestions for location of new rules
New HTE for Model Discrimination via Experimentation
Experiments in new composition regimesMeasure critical variables identifiedHigh Throughput
Experiments
time
Rat
e/S
elec
tivity
Performance Curves
FTIRGC...
Chemistry RulesReactionsLumping
Chemistry RulesReactionsLumping
Feature Extractor/Analyzer
Feature Extractor/Analyzer
Inverse Model
Genetic Algorithms
SelectionRecombination
Catalyst Library
HTE
Target Catalyst
Model Revision
Compare Performance
Catalyst Design Challenge
Target Catalyst Performancetime
Rat
e/Se
lect
ivity
Forward Model
Hybrid Model
Physical ModelStatistics/Neural-Nets
A + S A-Sk1
C + R Dk2
A-S +Dk3
time
Rat
e/Se
lect
ivity
Kinetics AI/Systems Tools Catalyst Performance
F
Pseudo Global Optimizer
Pseudo English Rule Compiler
Statistical Analyzer
Feature Extractor
Catalyst
{k}
Reaction Modeling Suite
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
LUBRIZOL: Fuel Additive Design
EPA requirement: Minimize intake-valve deposits (IVD)Approach: Fuel Additives Performance measure
BMW Test for IVD Stipulated to be less than 100 mg over a 10,000 mile road test
Expensive and time-consuming testing
Around $8000 for a single datum
Problem: Design fuel-additives that meet desired IVD performance levels
Intake Valve and Manifold
Fuel + Oil (Liquid + Vapor Mixture)
Additive Componentthat keeps it in solution
Additive Component that bindsto deposit “precursors”
Fuel-Additive Molecules
High-Boiling Deposit Forming Precursors
Fuel-Additives : A Functional Description
Component “length” directlyindicative of stability of additive
Breakage of this bond removes “dirt” carrying capacity totally
Chemical nature of this component (polar/non-polar) controls “dirt” removingcapacity
Breaking of these bonds control “length”
The first-principle model tracks the structural distribution of fuel-additive with time due to reactive degradation
FirstFirst--Principles Model for Additive DegradationPrinciples Model for Additive Degradation
Performance
Pattern RecognitionNeural Networks
RegressionAb Initio Methods Group Contribution
TopologicalTechniques
MolecularMechanics
Material/MoleculeFormulation
Primary Model
Structural Descriptors
Secondary Model
General Framework of Predictor ModelsGeneral Framework of Predictor Models
Modeling DegradationModeling Degradation
Dynamic in natureModeled as first-order irreversible reactions for additive degradationRate constants in proportion to rates of pure thermal degradation
Distribution of molecular species differing in structureIdentified by effective "length"Population balance model tracks distribution with time
From Stability to IVDFrom Stability to IVD
Define "amount of active additive"The fraction of the additive species that remains active (i.e.
intact viable structure) in the fuel at any point of time
A dynamic quantity decreasing with timeDepends on additive distribution
Solubility correlated to IVD via regressionLinear modelsNeural-Network modelsInput descriptors are the amounts of active additive at
different times
If A goes higher, B should go higher .....If C goes higher, D should go lower .....
Expert Rules/ HeuristicsFundamental Physics/Chemistry
)v(t
ρ•∇−=∂ρ∂
)(P)v(vt
)v(τ•∇+∇−=ρ∇•+
∂ρ∂
How to integrate diverse scientific knowledge/information into
a single unified knowledge architecture ?
Forward Problem Approaches
Empirically speaking, a polynomial expansion fits my data............
Data-driven/empirical models
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
Modeling Philosophy
All models are wrong….… but some are useful!
-- George Box(U. Wisconsin)
Hybrid ModelsCombine First-principles and Data-driven models
Hybrid Model for IVD PredictionHybrid Model for IVD Prediction
Time
Am
ount
of A
ctiv
e A
dditi
ve
τ0 τ2 τi τN
Additive A
Additive B
Intake-Valve Deposit
ComputerComputer--Aided Design of FuelAided Design of Fuel--AdditivesAdditives
Hybrid Model
Additive Performance
Intake ValveDeposit
Additive StructureFuel PropertyEngine Conditions
Genetic Algorithms
.
.
SelectionRecombination
Physical Model
Statistical/Neural-Net Correlation
Previous Approaches to Inverse Problem
Previous MethodologiesRandom SearchHeuristic EnumerationMath ProgrammingKnowledge-Based SystemsGraph Reconstruction
DisadvantagesCombinatorial Complexity
Nonlinear Search Spaces Local Minima Traps
Difficulties in Knowledge AcquisitionDifficulties in using high-level chemical/bio-chemical knowledge
Overview of Genetic AlgorithmsOverview of Genetic Algorithms
DefinitionGenetic Algorithms are stochastic, evolutionary search procedures based on Darwinian model of natural selection
Essential ComponentsGenetic Operators
CrossoverMutation
Reproductive PlanFitness Proportionate Selection
Genetic Algorithms for Product DesignGenetic Algorithms for Product Design
Global SearchDiversity of solutionsHigh potential for noveltyGlobal Optima
Development is de-coupled from forward problemRobust to non-linearity
Population based searchAbility to provide several near-optimal solutions
Captures transparently the rich chemistry of the design problem
Overview of Genetic AlgorithmsOverview of Genetic Algorithms
Select Operator
Molecular DesignsInitial Population
Calculate Fitness
Select Parent(s)
MutationCrossover
Apply Operator
Create New Population
Genetic Algorithms (GA)Genetic Algorithms (GA)
GAs are stochastic evolutionary search procedures based on the Darwinian model of natural selection
Initial Population (random)
Fitness Calculn, Parent Selectn
OperatorsNew
Population
“ Survival of the fittest ”
Evolution
Single-point Crossover
Genetic Operators: CrossoverGenetic Operators: Crossover
]nO C
OCO[O
H
H
C
H
H
C
O
H
H
C
H
H
C
CH3
CH3
C O CO[ ]n
O O CO
CO]n[
CH3
CH3
C O CO
O ]n[Parent #1 Parent #2
Offspring #1 Offspring #2
Main-chain and Side-chain Mutation
Mainchain MutationCH
HCH
H
n
CH
H
Parent: Offspring:
CH
H
n
CH
H
CH
HCF
H
n
Parent:
CH
HCH
n
Offspring:
Sidechain Mutation
Replace F by
Replace CH 2by
Mutation OperatorsMutation Operators
Fitness Function for GA-based CAMD
α = 0.0001
α = 0.01α = 0.001
α = 0.0005
Fitness (F) = exp - α Pi – PPi,max – Pi,min
2
Fractional Error (FE) = Pi – P
P
Polymer Design Case Study
• Base Groups
• Target Properties•
••
•
• Density Glass Transition Temperature
Thermal Expansion Coefficient Specific Heat Capacity
Bulk Modulus
15 Side-chain Groups
3 -C2H5 -nC H3 7 -iC H3 7
-Cl -Br -OH
-H
tC4H9
-OCH3-CN-
-CH
-F
3-C-O
O CH - 3-C-O
CH --C-O
NH--NH
X
X
X
17 Main-chain Groups>C< -S- -SO2- -O- -NH-
O-C-O
-C-O
-C-O
-O O--C-O
O--C-O
-C-O
NH- -O-C-O
NH-
Target Properties
Target PolymerDensity (gm/cm3)
2.96x10-4 1152.67 5.18x109
TP2: 1377.82 2.51x109
TP3: 1135.10 5.40x109
1073.96 5.39x109
TP5: 995.95 5.31x109
TP6: 1016.55 6.12x109
TP7:
1.34
1.18
1.21
1.19
1.28
1.25
1.06
350.75
225.24
420.83
406.83
472.00
421.12
322.55 1455.90 3.85x109
C
O
O C
O
O C
H
H
C
H
H n
C
H
H
C
F
F
C
H
H
C
H
CH 3 n
C
O
OO
CH 3
C
CH 3 n
CH
S O3
C
CH 3
O
n
O
n
SO 2 O
O O C
O n
C
H
H
C
H
H
C
H
H
C
H
H
C
H
H
C
O n
N
H
Glass Transition
(K)
Thermal Expansion
1/K
HeatCapacityJ/kg K
BulkModulusK N/m3
TP1:
2.81x10-4
2.90x10-4
2.90x10-4
2.90x10-4
2.98x10-4
2.89x10-4
TP4:
Polymer Design Case Study: Results
Legend: Success Rate; Average Conv. Generation; Number with fitness >0.99
Target Polymer
C
O
O C
O
O C
H
H
C
H
H n
C
H
H
C
F
F
C
H
H
C
H
CH 3n
C
O
OO
CH 3
C
CH 3 n
S O
CH 3
C
CH 3
O
n
O
n
SO 2 O
O O C
O n
C
H
H
C
H
H
C
H
H
C
H
H
C
H
H
C
O n
N
H
C
O
O C
O
O C
H
H
C
H
H n
CH 3
CH 3
O
CH 3
O
n
60%240213
48%5226
12%19374
0%-
589
48%232142
32%632168
100%64198
88%109161
4%86870
12%184282
36%4116
0%-
163
0%-
861
32%400175
8%548199
100%61217
68%210162
8%382144
Random MC, SC Random MC, SCFeasibility Constraints
Near Optimal DesignsPolymer Design Overall Error Fitness
Target Polymer: #4
3
S On
CC H 3
C HO 0% 1.0
Case 1: Standard GA
0.74% 0.995n
H HCH
O CO
CH
n
2C H 5 OC
C NO O O C O 1.18% 0.991
Case 2: Modified GA, StabilityHO H O
0.25% 0.999nC H 3
CH
C OO C
1.10% 0.991
Case 3: Modified GA, Stability & Complexity
O C HSC O C
2 5
H n
O CO
CH
H n
0.21% 0.999
HC 3 O O
nC
C H 3O C O C 0.83% 0.995
Advantages of GAs for Product Design
Global SearchDiversity of solutionsHigh potential for noveltyGlobal Optima
Development is de-coupled from forward problemRobust to non-linearity
Population based searchAbility to provide several near-optimal solutions
Captures transparently the rich chemistry of the design problem
•••
Powerful generic method with some drawbacksNo convergence guaranteesPerformance sensitive to parametersPerformance dependent on search space structure
Drawbacks of GAs
Evolutionary Design of FuelEvolutionary Design of Fuel--AdditivesAdditives
• Knowledge Based
Initial Population• Random
No
Termination Criteria• Last Generation ?• Fitness = 1.0 ?
New Generation
ProbabilisticSelect Operator
Select Parent(s)Fitness Proportionate
Apply Operator
Crossover Mutation
Elitist PolicyRetain k best in population
Calculate Fitness (F)Structure ---> Hybrid Model
Fitness <---- Additive Performance
Head LinkerBranch Tail
Yes
Stop
StartIVDdesired< IVDlimit;Maximize Solubility
Objective
Branch Crossover
L
JSite 2
Branch A
Parent-1
N
K
Parent-2
L
N
K
Branch A moved to site-2 to ensure feasibility of offspring-1
Offspring-1
J
Offspring-2
Genetic Operators: Branch CrossoverGenetic Operators: Branch Crossover
Run Rank/Identifier Fitness Predicted IVD(PLS-NN Model) Structural Description
1, I-1 0.997 11.4 mg Novel Structure. Infrequently used linker.
2, I-2 0.996 11.5 mg Novel Structure. Same tails as beststructure, different heads and linkersI
6, I-6 0.993 12.0 mgVariant of structure found in the BMWdatabase. Same head and linkers, differenttails
1, II-1 0.999 10.1 mg Novel Structure. Different from I-1.Infrequently used linker component.
2, II-2 0.989 12.6 mgSlight variant of additive structure foundin BMW and HONDA databases.Different tails but same head and linker
II
4, II-4 0.983 13.2 mg Minor variation of structure II-2 above.Slight modification of the head
1, III-1 1.00 8.9 mg Novel Structure. Different from I-1 andII-1. Commonly used components
2, III-2 0.994 11.9 mg Variant of III-1. One linker and tailmodified.III
3, III-3 0.993 12.1 mgVariant of structure II-2 above. Slightmodification of head. A linker and tailinserted.
Evolutionary Design of Fuel Additives: ResultsEvolutionary Design of Fuel Additives: ResultsObjective: Determine a structure with IVD < 10 mgDosage: 50 PTB; Population Size: 25; Generations: 25Objective: Determine a structure with IVD < 10 mgDosage: 50 PTB; Population Size: 25; Generations: 25
School of Chemical Engineering, Purdue University
Tires, Treads, Hoses, Shock Absorbers, O-rings, Gaskets, Mounts ……..
Rubber Parts in Service
School of Chemical Engineering, Purdue University
Rubber Formulation
Accelerator Activator
Rubber
Retarder
Filler
Antidegradants Oils
Sulfur
Mixing (~ hrs)
Part
Curing (~ hrs) Therm
al History
Final Vulcanized Part
Chemical Environment
Deformational Environment
Thermal Environment
Application (~ yrs)
Given the Desired Property/Performance of
Interest, what is the Optimal Rubber Formulation?
Problem Definition
Rubber Formulation
Accelerator Activator
Rubber
Retarder
Filler
AntidegradantsOilsSulfur
Which materials to include?
How much of them to include?
What should be the processing conditions?
School of Chemical Engineering, Purdue University
What are Sulfur-Vulcanized Elastomers?
Rubber Formulation
Accelerator Activator
Rubber
Retarder
Filler
AntidegradantsOilsSulfur
Accelerated Sulfur Vulcanization
SS S
S
Vulcanization
+ S8
School of Chemical Engineering, Purdue UniversityAIChE Annual Meeting, Reno 2001
OBJECTIVE OF THIS WORK
• Understand the Mechanistic Details of Accelerated-Sulfur-Vulcanization.
• Provide a Mathematical Description of the Kinetics of Accelerated-Sulfur Vulcanization consistent with Mechanistic Chemistry
Rubber Formulation
Accelerator Activator Retarder
Rubber Filler
Sulfur Oils Antidegradants
NC
SS N O
NR
ZnO/Stearic-Acid
BenzothiazoleSulfenamide
S8
School of Chemical Engineering, Purdue UniversityAIChE Annual Meeting, Reno 2001
COMPLEXITY OF VULCANIZATION REACTIONS
W.Scheele (1956) 1 Perhaps no-where in chemistry is there encountered a field which even in its literature alone shows so many uncertainties and (possibly only apparent) contradictions as that of the vulcanization of rubber.
L.Bateman (1963) 2 Whilst it has long been appreciated, albeit intuitively, that sulfur vulcanization is a very complex chemical process, the actual complexity as revealed in the studies described above is probably far in excess of what has ever been envisaged.
1W. Scheele, O. Lorenz and W. Dummer, Rubber Chemistry and Technology, 29, 1 (1956)
2 L. Bateman, C.G. Moore, M. Porter and B. Saville, “The Chemistry and Physics of Rubber-Like Substances”, L Bateman, Ed., Maclaren & Sons Ltd., London, 1963, pp 449.
P.J. Nieuwenhuizen (1997) 3 It must be considered remarkable that despite all the effortsand progress in the field of vulcanization during the past decade, one has to conclude that the statements of Scheele and Bateman, made 30 to 40 years ago, are still true to a great extent.
3 P.J. Nieuwenhuizen, Rubber Reviews, Rubber Chemistry and Technology, 29, 70 (1997)
School of Chemical Engineering, Purdue UniversityAIChE Annual Meeting, Reno 2001
SCHEMATIC OF A VULCANIZATION CURVE
Temporal Evolution of Crosslinks
Time
0
Con
cent
ratio
n of
Cro
sslin
ks, ν
Curing Phase
Reversion Phase
Induction Phase
A SS
SS
S
S S S
Time Scale ~YearsTime Scale ~ Hrs
School of Chemical Engineering, Purdue University
CAMD Sub-Problems
Structure Space
Pro
perty
INVERSE PROBLEM(Given Properties, Predict
Structure/Formulation)
Properties/ Performance( P1, P2, ……..PN )
FORWARD PROBLEM(Given Structure, PredictProperties/Performance)
Material Structure/Formulation
NeuralNetwork
Hybrid Models
Fundamental Models
School of Chemical Engineering, Purdue University
THE
OV
ERA
LL F
OR
WA
RD
MO
DEL
SubAssemblyDesignFull
ProductDesign &Operation
Design &Operation ofComplexSub-Assembly
Engine
Over 1000 rubber parts infailure critical functions
Mixing+
Heat Transfer+
Mold Filling
Part Manufacturing
Overcure
Undercure
Manufac-turing
Process
MaterialFormulation
Ax + S8k1⎯ →⎯⎯ Ax+8
Ax + R k2⎯ →⎯⎯ Bx
Bxk3⎯ →⎯⎯ Bx
*
etc.Time
Torque Cure Curve
Kinetics Polymer Network Mechanical Response
T̃ = 2ρR∂ ψ∂I 1
+ I 1∂ ψ∂I 2
⎧⎨⎩
⎫⎬⎭
I −∂ ψ∂I 2
C + I 3∂ ψ∂I 3
C−1⎡
⎣⎢⎤
⎦⎥••∂ C∂C
Constitutive Model
3D FiniteElementPart Model
Spring-DashpotApproximationwith ChemicalAging
PartShapeDesign
QuantumChemistry
SC
N S SS
SS
S
S
S
Zn
NewMolecules
School of Chemical Engineering, Purdue University
Accelerator Activator+
Active Accelerator Complex
S8
Active Sulfurating Agent
Rubber Bound Intermediate(Crosslink Precursor)
Rubber Bound persulfenyl radical
Polysulfidic crosslinks
Aged Network
DesulfurationDegradation
NC
SS Sx S C
S
NAx
Bx
C
C
CH Sx S CS
N
Bx•
C
C
CH Sx*
Vux
C
C
CH Sx C
C
C
NC
SS N O (ZnO + Stearic Acid)
2-morpholinothiobenzothiazole
Schematic of a Vulcanization Curve
School of Chemical Engineering, Purdue University
Accelerator Activator+
Active Accelerator Complex
S8
Active Sulfurating Agent
NC
SBt =
BtS-NR2 BtSH + R2NH
BtS-NR2 + BtSH BtSSBt + R2NH
BtS-SBt + S8 BtS-S-SBt + S7 BtS-S2-SBt + S6 •••
BtS-Sx-SBt + BtS-Sy-SBt BtS-Sz-SBt + BtS-Sw-SBt
NC
SS Sx S C
S
N
NC
SS N O
NC
SS Sx S
NC
SZn
L
L
Without Activator With Activator
Ax = BtS-Sx-SBt
Accelerator Chemistry
BtS-SBt + S8 BtS-S8-SBt
School of Chemical Engineering, Purdue University
CROSSLINKING CHEMISTRY (I)
Rubber Bound Intermediate(Crosslink Precursor)
Active Sulfurating AgentN
CS
S Sx S CS
N
CH
HC
C
CH
3
H
S
N C
S
S
N C
S
x S
S
CH
C CH
S
S
x
N
C
S S
+
N C
S
S
H
N C
S
SH
(R.34)
C
C
CH Sx S CS
N
Ax Bx
Rubber
Crosslinking Chemistry (I)
School of Chemical Engineering, Purdue University
Rubber, RH
BtS•
S8S8
S7RS-Sy+1
•
RS-Sy+8•
S8
S8
S7Bt Sz+1•S8
S8
(Regene ra tio n o f
S ul fu rat in g Sp ec ie s)
Bt Sz+8•
+ Bt-SH
+ Bt-Sz•
Crosslink PrecursorRSx –SBt, RSx-Zn-SBt
Crosslink PrecursorRSx –SBt, RSx-Zn-SBt
CrosslinksRSSy -RCrosslinksRSSy -R
Rubber, RH
Sulfurating SpeciesBtS-Sx-SBt, BtS-Zn-Sx-SBt
Sulfurating SpeciesBtS-Sx-SBt, BtS-Zn-Sx-SBt
Persulfenyl RadicalRS-Sy
•Persulfenyl RadicalRS-Sy
•
Rubber, RH Bt Sz+2•
Rubber, RH
RS-Sy
Cyclic SulfideBtSZnSBt
Main- Chain Modification, Conjugated Dienes/Trienes, Inactive thiols
RSSy-1R + BtSZnSSBt
(Desulfuration) (Degradation)
ZnOBtSZnSBt
BtS-ZnSx-SBt
(Scorch Delay)
CDB + Pthalimide (Retarder Action)
S8… S7 S8BtS-SBt
BtSNR2
BtSSSBtBtSSxSBt BtSS8SBtBtS-SBt BtSSxSBt
+ BtSSySBt
BtS-SBtBtSSxSBt + BtSSySBt
+ CTP-R2NH
Lumped S8 Pickup
Sequential S8 Pickup
Both
Mechanisms
BtSH 820 Reactions, 107 Species
School of Chemical Engineering, Purdue University
Equations for MBTS and other sulfurating species
[ ]
{ } { }{ } { }{ }{ }{ }{ }{ }{ }{ }⎥
⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢
⎣
⎡
++++++
++++++++++++
++++++++++
++++++++
++++++
++++++
+−+−−
−−−−=
∑∑
∑
=
∗
=
∗
=
278695104113122131
768594103112121
26758493102111
65748392101
2564738291
54637281
24536271
4352612
34251
32412
231212
1
A-A
16
2rr0DESULF
20E-E
16
1rr0BST-A0R-A
80S-A
14
2rr0A-AMBT-MBSMBS0
]A[5.0]A][A[]A][[A]A][[A ]][A[A]][A[A]][A[A2
]A][A[]A][[A]A][[A ]][A[A]][A[A]][A[A2]A[5.0]A][[A]A][[A ]][A[A]][A[A]][A[A2
]A][A[]A][[A ]][A[A]][A[A]][A[A2]A[5.0]A][[A ]][A[A]][A[A]][A[A2
]A][[A ]][A[A]][A[A]][A[A2]0.5[A ]][A[A]][A[A]][A[A2
]][A[A]][A[A]][A[A2][A5.0]][A[A]][A[A2
]][A[A]][A[A2][A5.0]][A[A2 ]][A[A2][A
k
Vu][Ak][Ek.50B][Ak][Ak
][S][AkA1)(r][Ak[MBS][MBT]k[MBS]kAdtd
][E][EkA][A2k
Vu][AkA][A2k][S][Ak][A2k][Adtd
10E-E
13
1rr1A-A
16
2rr0DESULF
14
2rr0A-A81S-A1R-A1
∗∗
=
==
+−
++−−=
∑
∑∑
POPULATION BALANCE EQUATIONS
School of Chemical Engineering, Purdue University
POPULATION BALANCE EQUATIONS
[ ] ⎥⎦
⎤⎢⎣
⎡+−−−= ∑
=
16
1ii0R114A81A31C11 Vu][Ak]A[]MBTS[k2]S[][Ak][Ak2A
dtd
[ ] [ ][ ] ][AkA1)(i][Ak][S][AkMBSMBTkAdtd
0C1
14
2ii0A480A3A20 −⎥⎦
⎤⎢⎣
⎡−−−= ∑
=
⎥⎦
⎤⎢⎣
⎡−⎥
⎦
⎤⎢⎣
⎡− ∑∑
==
16
1ii01R
16
1ii0C4 Vu]A[k*B][Ak
[ ]⎩⎨⎧
>≤≤
++−⎩⎨⎧
>≤≤
−−=7i]S[6i2,0
]A[k]A[][Ak1)(i7i,0
6i2],S[]A[k ][Ak2A
dtd
8i3Ai0A4
8i3AiC1i
Equations for MBTS and other sulfurating species
Equations for Crosslink Precursors
[ ] )*][Ek][MBTSk(Bdtd
1C7C10 +=
[ ] *][Ek]B[k)1i(][Ak2*]B][A[kBdtd
1iC7i2CiC1i0A4i ++−−+−=
School of Chemical Engineering, Purdue University
[ ]
[ ]
jconcanditimeatPointDataalExperiment
ConstantsRate
cdtdts
kkkk
ji
P
q
i
m
j ji
jiji
=
=≥
=
=⎥⎥
⎦
⎤
⎢⎢
⎣
⎡ −∑∑= =
exp,
321
2
1 1exp,
exp,,
k
k0k
)k,,(
..
......,,k,)k(
min
ν
νφν
ν
νν
m time points, q concentrations
RATE-CONSTANT DETERMINATION
FACTS
Total of 820 different reactions 107 Coupled Ordinary Nonlinear Differential Equations9 Optimizable Rate Constants
School of Chemical Engineering, Purdue University
B2•
Example Writing Population Balance Equations
kB
C
C
HC S S C
S
N
S
S * C
S
N
+
C
C
CH S 2*
+
C
C
CH S*
* C
S
N
SS
kA
B2
E•
E2•
Potential Breaking Sites
B•
[ ] ( )[ ]
[ ] [ ] [ ] [ ]
[ ] [ ] [ ] [ ]2B22B
2A2A2
B0BB
A0AA2BA2
BkEdtd,BkB
dtd
BkEdtd,BkB
dtd
RTEexpkk,
RTEexpkk,BkkB
dtd
==
==
⎟⎠⎞
⎜⎝⎛−=⎟
⎠⎞
⎜⎝⎛−=+−=
••
••
School of Chemical Engineering, Purdue University
Equations for MBTS and other sulfurating species
Population Balance Equations
[ ]
y
13
1x
x13
1yxA-A
16
2rr0DESULF
20E-E
16
1rr0BST-A0R-A
8
1yy0S-A
14
2rr0A-AMBTMBS-0
AAk
Vu][Ak][Ek21B][Ak ][Ak
][S][AkA1)(r][Ak[MBS][MBT]kAdtd
∑∑
∑∑
∑∑
=
−
=
=
∗
=
∗
==
+
−+−−
−−−=
][E][EkA][Ak2
Vu][AkA][A2k][S)][A - ][A(k][A2k][Adtd
10E-E
13
1rr1A-A
16
2rr0DESULF
14
2rr0A-A
8
1 y y01S-A1R-A1
∗∗
=
===
+−
++−−=
∑
∑∑∑
]A][[Ak
][E][EkA][Ak2]][A[Ak1)(x
A][A2k][S)][A- ][A(k][A2k][Adtd
r-x
1-x
1rrA-A
x0E-E
x14
1rrxA-Ax0A-A
14
1xrr0A-A
8
1yy1-xxS-AxR-Ax
∑
∑
∑∑
=
∗∗−
=
+==
+
+−−−
+−−=
x = 0
x = 1
2 ≤ x ≤ 13
School of Chemical Engineering, Purdue University
Final Set of Population Balance Equations
( )
( )
( )
( )
( )( )
( )⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢
⎣
⎡
=
⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢
⎣
⎡
∗
∗
∗
∗
∗
EBxBxAxkkkkfAxkf
SAkf
VuBxkkf
BxBxAxkkkf
BxBxAxkkkf
Axkkkkf
EMBT
S
Vu
Bx
Bx
Ax
xx
,,,,,,,,
,,...
,,,...
,,,,,...
,,,,,...
,,,,...
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5251427
16
80015
634
4323
4212
4102011
8
ddt Initial Conditions:
Ax(>0)(0) = Bx(0) = B*x(0) = Vux(0) = MBT(0) = E(0) = 0
A0(0) = [MBS]0 S8(0) = [Sulfur]0
Most simple kinetic model that explains all the important features of sulfur vulcanization.
Total of 820 different reactions considered with 107 coupled ordinary nonlinear differential equations and 9 optimizable rate-constants.
School of Chemical Engineering, Purdue University
Population Balance Model - Predictions
Open Symbols = Experimental DataSolid Line = Model PredictionsOpen Symbols = Experimental DataSolid Line = Model Predictions
T = 330 FT = 330 F
Effect of Accelerator and Sulfur
(1.0, 4.0)(0.5, 4.0)
(0.5, 2.0)
+ (1.0, 2.0)
(0.75, 2.0)x
Legend
School of Chemical Engineering, Purdue University
Population Balance Model – Different Temperatures
T = 310 FT = 310 F T = 298 FT = 298 F
Open Symbols = Experimental DataSolid Line = Model PredictionsOpen Symbols = Experimental DataSolid Line = Model Predictions
School of Chemical Engineering, Purdue University
QuantumChemistry
T̃ = 2ρR∂ ψ∂I 1
+ I 1∂ ψ∂I 2
⎧⎨⎩
⎫⎬⎭
I −∂ ψ∂I 2
C + I 3∂ ψ∂I 3
C−1⎡
⎣⎢⎤
⎦⎥••∂ C∂C
Constitutive Model
3D FiniteElementPart Model
Spring-DashpotApproximationwith ChemicalAging
Ax + S8k1⎯ →⎯⎯ Ax+8
Ax + R k2⎯ →⎯⎯ Bx
Bxk3⎯ →⎯⎯ Bx
*
etc.Time
To rq u eCure Curve
Kinetics Polymer Network Mechanical Response
Mixing+
Heat Transfer+
Mold Filling
Part Manufacturing
Overcure
Undercure
SC
N S SS
SS
S
S
S
Zn
NewMolecules
MaterialFormulation
Manufac-turing
Process
PartShapeDesign
SubAssemblyDesignFull
ProductDesign &Operation
Design &Operation ofComplexSub-Assembly
Engine
Over 1000 rubber parts infailure critical functions
THE
CA
TER
PILL
AR
GR
AN
D C
HA
LLEN
GE
School of Chemical Engineering, Purdue University
Spatial Profile of the State-of-Cure for theThermal History in the Mold
T he r
mal P
r of ile
85-87
89-91
91-93
93-95
95-97
> 99
97-99
87-89
Cur e
Pro
file
Actual Part Design
School of Chemical Engineering, Purdue University
Result Summary Vulcanization
Vulcanization Chemistry
MathematicalQuantification
Insights into Actual Part Design
Mechanistic Evaluation
• Evaluated all possiblemechanisms critically for accelerated sulfur vulcanization
• Identified the simplest set of reactions that describe all the important aspects of cure.
• Developed Population Balance Model based on mechanistic details to describe the different aspects of cure.
• Single set of parameters that predicts cure details for all formulations at all temperatures.
• Model extended for retarders, filled systems and other accelerator and elastomer classes.
• Incorporation of PopulationBalance model with a finiteelement code to predict spatial cure profile.
• Identification of undercuredand overcured regionsvisually.
School of Chemical Engineering, Purdue University
QuantumChemistry
T̃ = 2ρR∂ ψ∂I 1
+ I 1∂ ψ∂I 2
⎧⎨⎩
⎫⎬⎭
I −∂ ψ∂I 2
C + I 3∂ ψ∂I 3
C−1⎡
⎣⎢⎤
⎦⎥••∂ C∂C
Constitutive Model
3D FiniteElementPart Model
Spring-DashpotApproximationwith ChemicalAging
Ax + S8k1⎯ →⎯⎯ Ax+8
Ax + R k2⎯ →⎯⎯ Bx
Bxk3⎯ →⎯⎯ Bx
*
etc.Time
To rq u eCure Curve
Kinetics Polymer Network Mechanical Response
Mixing+
Heat Transfer+
Mold Filling
Part Manufacturing
Overcure
Undercure
SC
N S SS
SS
S
S
S
Zn
NewMolecules
MaterialFormulation
Manufac-turing
Process
PartShapeDesign
SubAssemblyDesignFull
ProductDesign &Operation
Design &Operation ofComplexSub-Assembly
Engine
Over 1000 rubber parts infailure critical functions
THE
OV
ERA
LL F
OR
WA
RD
MO
DEL
School of Chemical Engineering, Purdue University
CAMD Framework
The solution of INVERSE PROBLEM involves searching for the optimal rubber formulation that has the desired macroscopic performance.
Prediction
P1 P2 P3 ... Pn
Structure or Formulation Product Performance
Design
Inverse Problem
Prediction
P1 P2 P3 ... Pn
Structure or Formulation
Design
Inverse Problem
Forward Problem
RubberAcceleratorSulfurActivatorFillerRetarder
Cure-StateModulusStress-StrainStiffness
School of Chemical Engineering, Purdue University
Genetic Algorithms (GA)
GAs are stochastic evolutionary search procedures based on the Darwinian model of natural selection
Fitness Calculn, Parent Selectn
Fitn
ess
⎥⎥⎦
⎤
⎢⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛−
−−= ∑
2
mini,maxi,
desi,i
PPPP
(F)Fitness αexp
Initial Population (random)
Formulation = [1 4 0.1 30 330]
School of Chemical Engineering, Purdue University
Fitness Calculn, Parent Selectn
Fitn
ess
“ Survival of the fittest ”
New Population
Evolution
Genetic Algorithms (GA, contd)
Operators
School of Chemical Engineering, Purdue University
Formulation Representation
Binary Representation
Rubber Formulation
Accelerator Activator
Rubber
Retarder
Filler
AntidegradantsOilsSulfur
Rubber Formulation
Accelerator ActivatorAccelerator Activator
Rubber
Retarder
Rubber
Retarder
FillerFiller
AntidegradantsOilsSulfur AntidegradantsOilsSulfurSulfur
0 01 1 •• •1 0 01 1
Phr of Different Ingredients
School of Chemical Engineering, Purdue University
Genetic Operators
Parent 1 Parent 2
Offspring 1 Offspring 2
0 01 1 10 0 01 1 1 01 1 01 0 01 0
0 01 1 0 1 00 00 1 01 1 1 1 10 01
0 01 1 10 0 01 1 1 01 1 01 0 01 0
0 01 1 0 1 00 000 01 1 0 10 01 1 0 1 00 00 00 00 1 01 1 1 1 10 011 01 1 1 1 10 01 10 01
Formulation
Fitn
ess
Crossover
0 01 1 10 0 01 1
1 01 1 10 0 01 1
0 01 1 10 0 01 1
1 01 1 10 0 01 1
Mutation
School of Chemical Engineering, Purdue University
Tmax (time to reach max cure) = 16 min
Vumax (crosslink @ Tmax) = 71.5 mol/m3
Stress (100% Elongation) = 0.92 MPa
Stress (200% Elongation) = 1.65 MPa
Optimal FormulationsDesired Property
Inverse Problem (Results)
• Binary Representation
• Fitness Proportionate Selection
• Crossover Probability = 0.8, Mutation Probability = 0.2
• Population Size = 40, Number of Generations = 40, Elitism = 10%
[1.75 1.25 0.05 0 315]
[0.50 4.00 0.10 0 310]
[1.75 1.50 0.05 0 310]
[0.75 2.75 0.10 0 315]
[3.00 0.75 0.10 0 315]
[2.75 1.00 0.25 0 330]
0.9999
0.9999
0.9996
0.9994
0.9991
0.9990
15
16
17
13
17.5
10.5
69.92
71.57
75.75
67.88
65.34
70.46
0.91
0.92
0.97
0.88
0.85
0.93
1.63
1.65
1.75
1.58
1.52
1.67
Formulation Fitness Tmax Vumax σ100 σ200
Example 2
School of Chemical Engineering, Purdue University
[1.75 1.25 0.05 0 315]
[0.50 4.00 0.10 0 310]
[1.75 1.50 0.05 0 310]
[0.75 2.75 0.10 0 315]
[3.00 0.75 0.10 0 315]
[2.75 1.00 0.25 0 330]
0.9999
0.9999
0.9996
0.9994
0.9991
0.9990
15
16
17
13
17.5
10.5
69.92
71.57
75.75
67.88
65.34
70.46
0.91
0.92
0.97
0.88
0.85
0.93
1.63
1.65
1.75
1.58
1.52
1.67
Formulation Fitness Tmax Vumax σ100 σ200
0 10 20 30 40 50 600
10
20
30
40
50
60
70
80
1
2
3
Time, min
Con
cent
ratio
n of
Cro
sslin
ks, m
ol/m
3
New Formulations
BEST THREE FORMULATIONS
Solution Found in 24th generation.
i.e.only 960 out of a total of 131072 formulations were searched
~ 0.73 %
School of Chemical Engineering, Purdue University
Interactive Software Used At Caterpillar on a Daily Basis
10/11/2005
1
Pharmaceutical Product Development and Engineering
• Two ProductsDrugsDocuments
Preliminary product recipe
Drug is converted into ParticlesDrug Synthesis
Raw Chemicals
Formulation
Process Development
Preliminary process recipe
Scale/up -Tech Transfer
Adjusted process recipe
Manufacturing
A Delayed Product
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 2
LIPS
• Prescription drug recalls• 176 in 1998• 248 in 2001• 354 in 2002
• Schering-Plough Corp. recalled 59 million asthma inhalers in 1999 and 2000 -- unknown number were shipped empty.•Semiconductor Manufacture 6 σ•Chemicals 5 to 5.5 σ•Pharma 2.5 σ
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 3
LIPS
Pharma Industry’s Scale-up Challenge
engineering
pharmaceutical
4
Purdue Ontology for Pharmaceutical Engineering (POPE)
Information
Deci
sion
tree
s an
d he
uris
tics
Mathem
atical models
ModLABOntoMODEL
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 5
LIPS
Ontologies Developed So Far in Our ProjectEquipment ontology
Standards referenced: STEP, AP231, FIATECHGeneral recipe ontology
RecipeElement, UnitProcedure, Operation etc. Standards referenced: ISA S88, S95, OntoCAPE
Process safety ontologyDeviation, Cause, Consequence etc.
Material ontology for pharmaceutical product developmentFlow property, Angle_of_Fall, Carrs_Index etc.
Reaction mechanism ontologyMolecule, Atom, Bond, Reaction etc. Referenced: Chemistry Development Kit
Model ontologyGuideline ontology
Referenced: GuideLine Interchange Format
Venkat Venkatasubramanian, PSWC, April 2007, Amsterdam 6
LIPS
Material Property Ontology in Protégé
Venkat Venkatasubramanian, PSWC, April 2007, Amsterdam 7
LIPS
Experiment Ontology
Experiment Data Files
Experimental Methods Details
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 8
LIPS
Guideline Ontology for Dosage Form Formulation
Follows GLIF (Guideline Interchange Format), A standard ontology for clinical guidelines
9
Model Ontology
Venkat Venkatasubramanian, Plenary Lecture,Foundations of Computer-Aided Process Design, Princeton, July 2004.
SuperModel
Model Container Model Container
Model Model
hasModelContainers
hasModel hasModelhasModelIndex hasModelIndex
Equation
Partial Differential
Integral
Algebraic
DAE
Function Evaluation
Assumptions
Universal Constants Dependent Var Independent Var Model Parameters
Variable Link
Variable
Variable
Simple Variable
Array Variable
SizeOfListVariable
Link
hasInitialGuess
hasVar PointsToValue
PointsToValuehasVar
hasEqn
hasAssumptions
hasUnivConstants hasDepVar hasIndepVar hasModelParms
hasVar
hasSymbol
hasML
Text Input
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 11
LIPS
Integrated Model-based Decision SupportRecipe Process Information Repository
Material Form in XForms Recipe Network
Used by PHASuite
Used by Batches/Batch+
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 12
LIPS
Integrating with Mathematical Knowledge
User created model instance
Model Ontology Equations in MathML
• Variables linked through URI
Process Description (instances of Process Ontology)
Automatically generated Mathematica statements
Interface
Mathematica Notebook
Access ontology instances
Mathem
atica K
ernel
Rich ClientWeb Applications
13
Literature Experiments
Experiment Report Data
Intelligent Systems for PreformulationFormulationProcess Monitoring and Control
Tools for Organic Solids Mfg.Optimization, Simulation, Safety Analysis, Scheduling
Multiscale models for synthesis mehtods
Modeling and Analysis Crystalline Structure
Unit Operation modelsDEM/CFD/FEM
Process modelingFirst Principles Process Model
Organic composites:Physical, chemical, and powder
properties
Equipment: blender, tabletting press, roller compactor, fluid bed
Process:Crystallization, Granulation, Drying,
Size Reduction etc.
Knowledgerelationships
Systematic functionalization
Control methodology for spatial structure of organic composites
Material synthesis method development
Determine structure-function-performance relationships
Instruments
Cyberinfrastructure for Real-time Decision Support for Design, Control, and Optimization
Dissolution Problems
Spatial Structure
Process Monitoring and Control
Tools for Design and Manufacturing
Modeling
User
Ontological Informatics Infrastructure
Material Science
Experiments
Process
DEM
Experiment Report Physical PropertiesPowder Properties
eLabNotebookBlender
Literature Experiments
Experiment Report Data
Intelligent Systems for PreformulationFormulationProcess Monitoring and Control
Tools for Organic Solids Mfg.Optimization, Simulation, Safety Analysis, Scheduling
Multiscale models for synthesis mehtods
Modeling and Analysis Crystalline Structure
Unit Operation modelsDEM/CFD/FEM
Process modelingFirst Principles Process Model
Organic composites:Physical, chemical, and powder
properties
Equipment: blender, tabletting press, roller compactor, fluid bed
Process:Crystallization, Granulation, Drying,
Size Reduction etc.
Knowledgerelationships
Systematic functionalization
Control methodology for spatial structure of organic composites
Material synthesis method development
Determine structure-function-performance relationships
Instruments
Tools for Design and Manufacturing
Modeling
User
Ontological Informatics Infrastructure
Material Science
Experiments
Venkat Venkatasubramanian, Keynote Lecture,European Congress in Chem. Engg, Copenhagen, Sept 2007 14
LIPS
SummaryReviewed Modeling and Informatics Challenges in Molecular Products Design and EngineeringNeed for Cyberinfrastructure Concepts, Methods, and ToolsOntological approach for information and knowledge modeling
Beyond ERP, SAP, Oracle, Expert Systems etc.This is not just programming! Nor can it be done by CS folks alone!
POPE: Purdue Ontology for Pharmaceutical EngineeringConceptual foundation for next the generation, integrated, decision support tools environment
This is only a beginning….Miles to go before we sleep….Huge opportunities for Process/Product Systems Engineers
15
Thank You for Your Attention!
Any Questions?