Optimization and datamining for catalysts library design

1
Optimization and datamining for catalysts library
design

• Frédéric CLERC

2
Contents

• Introduction
• Catalysts
• Combinatorial optimization
• Concept of meta modeling
• Results
• Concept validation
• Tuning of algorithm
• Case studies
• Conclusion and Future works

3
Catalysts

• Car exhaust pipe

• Honeycomb structure

4
Catalysts

• Powder deposit
• Favor and select chemical reactions

• Search for efficient compounds
• Computational tool combinatorial optimization

5
Combinatorial optimization

• Populations evolve through iterative process
• Generations converge towards optimum
• Individual virtual catalyst

Y
X1
X2
6
Generic optimization process
Random initialization
Population
OPERATORS
Evaluation
False
True
Best individual
7
Algorithms

• Taboo search
• Simulated annealing
• Genetic algorithms
• Evolutionary strategies
• OptiCat

• Diversity of data treatments

8
Issues in catalysis

• Evaluation requires manual expertise
• Synthesis
• Test
• Cost of a catalyst
• Time
• Money
• Skepticism against black box optimization

?
9
Solution Meta modeling

• Optimization and dataminig
• Genetic operators
• Datamining operator
• Data storage
• Supervised learning

OPERATORS

• Cost reduction
• Opening the black box

10
Meta modeling
Random initialization
Estimation
Population
Use the statistical model for estimating the
efficiency of virtual catalysts
OPERATORS
Catalysts Choice
Choose promising catalysts among the estimated
ones
Use a statistical model for predicting the
efficiency of virtual catalysts
Evaluate the chosen catalysts and update
statistical model
Evaluation
Choose promising catalysts amon the estimated ones
Evolution Control
Evaluate the chosen catalysts
False
True
Best individual
11
Objectives in catalysis

• Reduced number of evaluations
• Population size ? 40
• Number of generations ? 10
• Catalyst complexity gt 1020
• Other means of catalytic evaluations
• Prove the efficiency of combinatorial approach

12
Meta modeling validation

• Use virtual surface response
• Validate the approach
• Use purpose-designed data
• Tune algorithmic parameters
• Use real data
• Appropriate datamining algorithm for context
• Obtain results

13
Artificial evaluation
14
Response Surface and Dimensions

• Three optima
• PtPd support Al2O3
• Cu support CeO2
• Au support TiO2

15
Learning algorithm

• Continuous space by parts
• Single performance response
• Simple and efficient

• Multi linear regression

16
Algorithm
Random initialization
Linear regression
Population
Elitism
Evaluation
False
True
Best individual
17
Algorithm efficiency measure
v4
v3
v2

• Stochastic algorithm multiple runs for
  statistical relevance
• Need for an efficiency measure
• Performance
• Reliability

v5
v1
18
Results

• Compare with classic methods

• Good compromise reliability-performance

19
Summary

• Artificial response surface
• Multilinear regression learning
• Compared with classic algorithms
• Other surfaces, experimental noise

• Meta modeling is reliable and performing

• Still valid with purpose designed data?
• How to tune parameters?

20
Meta modeling validation

• Use virtual surface response
• Validate the approach
• Use purpose-designed data
• Tune algorithmic parameters
• Use real data
• Appropriate datamining for context
• Obtain results

21
Purpose designed data

• 168 catalysts prepared for CO oxidation

• Four variables
• Noble metal Au, Cu, Pt
• Transition metal Mo, Nb, V
• Support CeO2, TiO2, ZrO2
• Reaction temperature 200, 250, 300C
• Performance function
• CO Conversion

22
Algorithm Tuning
Random initialization
Linear regression
Population
Selection Crossover Mutation
Elitism
Genetic operators
Evaluation
false
true
Best catalyst
23
Reduce algorithm number
4 Population sizes 4 Selection types 4 Crossover
types 2 Elitism modalities
4?4?4?2 128 algorithms
Use Design of Experiments

• Reduce to 16 algorithms
• Quantify the impact of each parameter

24
Results
Elitism Yes No Population size 8 16 24 48
Crossover type One point Three points Uniform
20 Uniform 50 Selection type Wheel Threshold
Tournament Rank
Low

• Average Importance
• Points-based crossover preferred

• Important
• Efficiency of tournament selection

Very High

• Modality weight
• Long bar gt important modality
• Elitism not important

• Very important
• High population size gt high efficiency

Average
High
25
Summary

• Use data-based surface CO oxidation
• Use DoE for efficient parameter tuning

• Meta modeling tuning realized
• High population size
• Tournament selection

• Which learning algorithm?

26
Meta modeling validation

• Use virtual surface response
• Validate the approach
• Use purpose-designed data
• Tune algorithmic parameters
• Use real data
• Appropriate learning algorithm with regards to
  context
• Obtain results

27
Search spaces
QSAR
Descriptor calculation
Elemental composition
Response space
Literature data
Optimization search space
Datamining search space
28
Case studies in catalysis

• Heck reaction
• 222 catalysts
• 30 continuous descriptors
• 3 performance values

• Propene Oxidation
• 467 catalysts
• 72 descriptors
• discrete
• continuous
• 1 desirability value

• Partial least squares regression

• Artificial neural network

29
Algorithm
Random Initialization
QSAR descriptor calculation
Population
Selection Crossover Mutation
Estimation
Genetic operators
Elitist choice
false
true
Best catalyst
30
Results

• Guidelines promising catalyst families

• Propene oxidation
• 14 Gallium
• 16 Niobium
• Support oxyde
• Solvant alcohol

• Heck
• parts of molecules

• Meta modeling appropriate for different contexts

31
Summary

• Two case studies in catalysis
• Propene oxidation ANN
• Heck reaction PLS
• Drawing guidelines for catalysts design

• Learning algorithm choice depends on context
• Type of variables
• Quantity of observations
• Classic datamining requirements

32
Conclusions

• Optimization and Datamining for catalysts library
  design
• Meta modeling reliable and performing
• Tuning general parameters
• Two case studies
• Fast catalysts optimization (400)
• Find promising guidelines
• OptiCat as multi-purpose
• software tool

33
OptiCat Diffusion

• Free (CeCILL license)
• http//eric.univ-lyon2.fr/fclerc
• On line model builder
• Webservice

WSDL
34
Future works

• Other learning algorithms
• Association rules
• Diversity control
• Multiple optima
• Real experimentation
• University of Amsterdam
• IRC (TOPCOMBI Program)
• Max Planck Institute

35
Acknowledgements
David Farrusseng Ricco Rakotomalala Ferdi
Schuth Gadi Rothenberg Claude Mirodatos Nicolas
Nicoloyannis Gilles Venturini Djamel
Zediar Silvia Pereira Enrico Burello Jos
Hageman Ignacio Lopez Martin Laurent
Baumes Mourad Lengliz Joanna Procelewska Javier
Llamas Galilea Juriaan Beckers Jan Blank
David Farrusseng Ricco Rakotomalala Ferdi
Schuth Gadi Rothenberg Claude Mirodatos Nicolas
Nicoloyannis Gilles Venturini Djamel
Zediar Silvia Pereira Enrico Burello Jos
Hageman Ignacio Lopez Martin Laurent
Baumes Mourad Lengliz Joanna Procelewska Javier
Llamas Galilea Juriaan Beckers Jan Blank
David Farrusseng Ricco Rakotomalala Ferdi
Schuth Gadi Rothenberg Claude Mirodatos Nicolas
Nicoloyannis Gilles Venturini Djamel
Zediar Silvia Pereira Enrico Burello Jos
Hageman Ignacio Lopez Martin Laurent
Baumes Mourad Lengliz Joanna Procelewska Javier
Llamas Galilea Juriaan Beckers Jan Blank
David Farrusseng Ricco Rakotomalala Ferdi
Schuth Gadi Rothenberg Claude Mirodatos Nicolas
Nicoloyannis Gilles Venturini Djamel
Zediar Silvia Pereira Enrico Burello Jos
Hageman Ignacio Lopez Martin Laurent
Baumes Mourad Lengliz Joanna Procelewska Javier
Llamas Galilea Juriaan Beckers Jan Blank
David Farrusseng Ricco Rakotomalala Ferdi
Schuth Gadi Rothenberg Claude Mirodatos Nicolas
Nicoloyannis Gilles Venturini Djamel
Zediar Silvia Pereira Enrico Burello Jos
Hageman Ignacio Lopez Martin Laurent
Baumes Mourad Lengliz Joanna Procelewska Javier
Llamas Galilea Juriaan Beckers Jan Blank

• David Farrusseng
• Ricco Rakotomalala
• Ferdi Schuth
• Gadi Rothenberg
• Claude Mirodatos
• Nicolas Nicoloyannis
• Gilles Venturini
• Djamel Zediar
• Silvia Pereira
• Enrico Burello
• Jos Hageman
• Ignacio Lopez Martin
• Laurent Baumes
• Mourad Lengliz
• Joanna Procelewska
• Javier Llamas Galilea
• Juriaan Beckers

36
Créer une population Opérateurs génétiques

• Le croisement mélange les caractéristiques des
  individus

• La mutation introduit de nouvelles informations

• La sélection retient les individus les plus
  adaptés

• Assurer une diversité contrôlée dans les
  populations

37
Sélection
30
90

• Deux ou plusieurs individus intègrent un tournoi

• Leur performance est comparée

• Seul le meilleur est admis pour létape suivante

• Il y a autant de tournois consécutifs que
  dindividus dans la population
• Des individus sont éliminés alors que dautres
  sont répétés

38
Croisement

• Un point de scission est déterminé dans les
  individus

• Croisement multi-points
• Croisement uniforme

• Les portions sont échangées

39
Mutation
70

• Introduit des changements aléatoires dans les
  valeurs

40
OptiCat
41
Meta modeling
42
Heck Reaction
43
Meta modeling