VALIDATION OF A RANDOM MATRIX MODEL FOR MESOSCALE ELASTICDESCRIPTION OF MATERIALS WITH MICROSTRUCTURES
ROGER GHANEM, University of Southern California, Los Angeles, CA
ARASH NOSHADRAVAN, Massachusetts Institute of Technology, Cambridge, MA
JOHANN GUILLEMINOT, Université Paris-Est, MSME, Marne-la-Vallée, France
IKSHWAKU ATODARIA, PEDRO PERALTA, Arizona State University, Tempe, AZ
April 5, 2012SIAM-UQ 2012 - Raleigh, NC
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 1 / 18
Introduction and motivation
Structural Health Monitoring (SHM) and Prognosis in Aerospace Systems
Future of SHM and Prognosis
Advances in experimentalcharacterization, sensing technology,computational facility.
human inspection, timed maintenance→ condition-based SHM, residual lifeestimation.
Anticipate damage from measureddata.
- number, type and location ofsensors
Sensing and Processing
Predictive modeling
Data Interrogation
Damage Prognosis
Structural Health Monitoring
Uncertainty Quantification & Model Validation
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 2 / 18
Challenges-Solutions
Challenge
Damage initiates at a very small scale.
Measured data is at a coarse scale.
The details of the microscale for the specimenbeing measured are not known.
Microscale Simulation ?microstructure unknown - can be characterized statistically in the lab.
to determine location of sensors we need to formulate an optimization problem.
mechanistic analysis of an ensemble of microstructures is very expensive.
SolutionDevelop a new stochastic mechanistic model with :
State of the model at same scale as experimental observables.
Model behavior sensitive to occurrences at the scale of damage initiation.
Scatter in predictions from model consistent with observed scatter.
Behavior of model honors known accepted conservation laws.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 3 / 18
Challenges-Solutions
Challenge
Damage initiates at a very small scale.
Measured data is at a coarse scale.
The details of the microscale for the specimenbeing measured are not known.
Microscale Simulation ?microstructure unknown - can be characterized statistically in the lab.
to determine location of sensors we need to formulate an optimization problem.
mechanistic analysis of an ensemble of microstructures is very expensive.
SolutionDevelop a new stochastic mechanistic model with :
State of the model at same scale as experimental observables.
Model behavior sensitive to occurrences at the scale of damage initiation.
Scatter in predictions from model consistent with observed scatter.
Behavior of model honors known accepted conservation laws.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 3 / 18
Challenges-Solutions
Challenge
Damage initiates at a very small scale.
Measured data is at a coarse scale.
The details of the microscale for the specimenbeing measured are not known.
Microscale Simulation ?microstructure unknown - can be characterized statistically in the lab.
to determine location of sensors we need to formulate an optimization problem.
mechanistic analysis of an ensemble of microstructures is very expensive.
SolutionDevelop a new stochastic mechanistic model with :
State of the model at same scale as experimental observables.
Model behavior sensitive to occurrences at the scale of damage initiation.
Scatter in predictions from model consistent with observed scatter.
Behavior of model honors known accepted conservation laws.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 3 / 18
Challenges-Solutions
Challenge
Damage initiates at a very small scale.
Measured data is at a coarse scale.
The details of the microscale for the specimenbeing measured are not known.
Microscale Simulation ?microstructure unknown - can be characterized statistically in the lab.
to determine location of sensors we need to formulate an optimization problem.
mechanistic analysis of an ensemble of microstructures is very expensive.
SolutionDevelop a new stochastic mechanistic model with :
State of the model at same scale as experimental observables.
Model behavior sensitive to occurrences at the scale of damage initiation.
Scatter in predictions from model consistent with observed scatter.
Behavior of model honors known accepted conservation laws.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 3 / 18
Challenges-Solutions
Challenge
Damage initiates at a very small scale.
Measured data is at a coarse scale.
The details of the microscale for the specimenbeing measured are not known.
Microscale Simulation ?microstructure unknown - can be characterized statistically in the lab.
to determine location of sensors we need to formulate an optimization problem.
mechanistic analysis of an ensemble of microstructures is very expensive.
SolutionDevelop a new stochastic mechanistic model with :
State of the model at same scale as experimental observables.
Model behavior sensitive to occurrences at the scale of damage initiation.
Scatter in predictions from model consistent with observed scatter.
Behavior of model honors known accepted conservation laws.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 3 / 18
Physically-based Multiscale Modeling for SHM and Prognosis
• Microscopic images
• Ultrasonic techniques
Probabilistic physically-based multiscale model
• Damage detection
• Life time evaluation
• Prediction & Prognosis
Characterization of microstructural heterogeneity
Stochastic characterization and representation of
microstructureModel Verification
& Validation Stochastic upscaling
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 4 / 18
Physically-based Multiscale Modeling for SHM and Prognosis
• Microscopic images
• Ultrasonic techniques
Probabilistic physically-based multiscale model
• Damage detection
• Life time evaluation
• Prediction & Prognosis
Characterization of microstructural heterogeneity
Stochastic characterization and representation of
microstructureModel Verification
& Validation Stochastic upscaling
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 4 / 18
Outline
1 Statistical model for characterization and realization of microstructure
2 Probabilistic model for mesoscale bounded elasticity tensor of polycrystalls
3 Validation methodology
4 Summary and conclusions
Statistical model for characterization and
realization of microstructure
Probabilistic model for mesoscale description
of materials
Model Calibration
Model Validation
Microstructural measurements
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 5 / 18
Experimental characterization and database
Three EBSD maps of 12 mm×1.5 mm Al-2024 specimenInformation on geometry and crystallographic orientation of ≈12000 grainsThe microstructure is described by the equivalent aspect ratios and Euler anglesof all the grains
0 1 2 3 4 5 6 70
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Euler angle (rad)
pd
f
φ1
φφ
2
pdfs of Euler angles
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.5
1
1.5
2
2.5
3
Aspect ratio β
pdf of overall aspect ratio
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 6 / 18
Statistical model for characterization and realization of microstructures
I : Random geometry
Voronoi-G tessellation
Extension of classical Voronoi-tessellating by using the following distance
V(x(i)tes)def= {x ∈ Ω | dG(x(i)tes, x) ≤ dG(x
(i)tes, x
(j)tes)},
dG(x, y)def=√
(x, y)T [G] (x, y)
[G] =[
(1/s)2 00 1
]
s : Rate of growth of tessellation in e1 − direction1/s controls the aspect ratio of the grains
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 7 / 18
Statistical model for characterization and realization of microstructuresMinimum Relative Entropy estimates of parameter s
ŝ = arg mins∈R+
DKL (p̂β(β) ‖ pβ(β|s)) ,
DKL (p̂β(β) ‖ pβ(β|s)) =∫R+
p̂β(β) logp̂β(β)
pβ(β|s)dβ.
DKL (Kullback-Leibler divergence) : distance-like measure of the difference between two pdfs
1 1.5 2 2.5 3 3.5 4 4.50
0.2
0.4
0.6
0.8
1
1.2
1.4
Parameter s
Rel
ativ
e en
trop
y D
KL
Plot of minimum relative entropy functions
−0.6 −0.4 −0.2 0 0.2 0.4 0.6
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
−0.6 −0.4 −0.2 0 0.2 0.4 0.6
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 8 / 18
Statistical model for characterization and realization of microstructures
II : Random crystallographic orientation
Objective
Euler angles are modeled as vector-valued random field.
x 7→ Φ(x) = (φ1(x), φ2(x), φ3(x))Approach
The vector-valued random field of Euler angles is modeled as a translationprocess with prescribing :
Marginal probability distribution functions estimated from measurementsFractile correlation functions estimated from measurements
Fractile correlation
um(x) = Fm(φm(x)), m = 1, 2, 3.
[R(x, x′)]mn =E{(um(x)− um(x))(un(x
′)− un(x′))}√
E{(um(x)− um(x))2}E{(un(x′)− un(x′))2}
= 12E{um(x)un(x′)} − 3.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 9 / 18
Statistical model for characterization and realization of microstructuresEstimation of correlation functions from measurements :
0 30 60 90 120 150−1
−0.75
−0.5
−0.25
0
0.25
0.5
0.75
1
η (micron)
[R] ii
[R]
11
[R]22
[R]33
0 30 60 90 120 150−1
−0.75
−0.5
−0.25
0
0.25
0.5
0.75
1
η (micron)
[R] ij
[R]
12
[R]23
[R]13
Identification of the correlation structure : [R(η)]mn = amn exp(−η/lmn)
0 30 60 90 120 150−1
−0.75
−0.5
−0.25
0
0.25
0.5
0.75
1
η (micron)
[R] 1
1
estimated valuesbest exponential fit (a
11=1.0, l
11=22.4553)
0 30 60 90 120 150−1
−0.75
−0.5
−0.25
0
0.25
0.5
0.75
1
η (micron)
[R] 1
3
estimated valuesbest exponential fit (a
13=−0.9416, l
13=35.9678)
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 10 / 18
Probabilistic model for mesoscale elasticity tensor
Objective
Mesoscale material description that :(i) captures the effect of subscale heterogeneities (ii) could be used in a coarse scalemodeling.
Approach : MaxEnt principle & Random matrix theory
Constructing a probability distribution on the set of elasticity matrices.
Constrain random matrices to specified physics-based bounds.
Calibrate the random matrices from all the available information.
Prob. model for mesoscale elasticity
tensor
Calibration
Verification Validation
Prediction & Detection
....
Microstructure
Macroscale (RVE) Mesoscale (SVE)
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 11 / 18
Probabilistic model for mesoscale elasticity tensor
Overview of model constructionLet :
N = (C− Cl )−1 − (Cu − Cl )−1 > 0,
Maximize : ∫M+n (R)
ln(p)pN(N) dN
subject to : ∫M+n (R)
pN(N) dN = 1,∫M+n (R)
N p[N](N) dN = N ∈ M+n (R),∫C
ln(det(N))pN(N)dN = cN , |cN | < +∞.
pN(N) = IM+n (R)(N)ĉ0 det(N)λ−1etr{−ΛN N}
ΛN and λ are Lagrange multipliers.
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 12 / 18
Probabilistic model for mesoscale elasticity tensorCalibration
Computing the realizations of the bounds for apparent elasticity matrix (Huet’s partitioningtechnique)Computing the realizations of the apparent elasticity matrix
Apparent properties
min ‖〈σ〉BC − [C]〈�〉BC‖subject to meaningful constraints
}⇒ [Capp]
BCs :
SUBC : t(x) = σ0n(x) ⇒ Lower bound [Cappσ ]
KUBC : u(x) = �0x ⇒ Upper bound [Capp� ]
MBC (Tension test, e.g.)⇒ Samples of apparent elasticity tensor [Capp]
Compute the statistical estimates of parameters for Nsim = 100 and Ω = 0.3× 0.3 [mm]
δ̃N =
1Nsim‖[Ñ]‖2FNsim∑k=1
‖[N(ωk )]− [Ñ]‖2F
1/2
= 0.66
[Ñ] =1
Nsim
Nsim∑k=1
[N(ωk )] = 10−3
0.2667 0.0879 −0.01890.0879 0.2214 0.0277−0.0189 0.0277 0.2366
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 13 / 18
Probabilistic model for mesoscale elasticity tensorCalibration
Computing the realizations of the bounds for apparent elasticity matrix (Huet’s partitioningtechnique)Computing the realizations of the apparent elasticity matrix
Apparent properties
min ‖〈σ〉BC − [C]〈�〉BC‖subject to meaningful constraints
}⇒ [Capp]
BCs :
SUBC : t(x) = σ0n(x) ⇒ Lower bound [Cappσ ]
KUBC : u(x) = �0x ⇒ Upper bound [Capp� ]
MBC (Tension test, e.g.)⇒ Samples of apparent elasticity tensor [Capp]
Compute the statistical estimates of parameters for Nsim = 100 and Ω = 0.3× 0.3 [mm]
δ̃N =
1Nsim‖[Ñ]‖2FNsim∑k=1
‖[N(ωk )]− [Ñ]‖2F
1/2
= 0.66
[Ñ] =1
Nsim
Nsim∑k=1
[N(ωk )] = 10−3
0.2667 0.0879 −0.01890.0879 0.2214 0.0277−0.0189 0.0277 0.2366
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 13 / 18
Probabilistic model for mesoscale elasticity tensor
VerificationWhether the model implementation accurately represent the intended conceptual descriptionof the model and the solution to the model
1.08 1.1 1.12 1.14 1.16
x 105
0
1
2
3
4
5
6x 10
−4
C11
1.08 1.1 1.12 1.14 1.16
x 105
0
1
2
3
4
5
6x 10
−4
C22
2.4 2.5 2.6 2.7 2.8 2.9
x 104
0
1
2
3
4
5
6
7
8x 10
−4
C33
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 14 / 18
Validation : Static
Credibility of the model in propagating the uncertainty in the fine scale into thecoarse scale response
x (mm)
y (
mm
)
0.0 1.0 2.0 3.0
0.0
1.0
2.0
3.0
0= 1000 N/mm2σ
Quantity of interest (QOI) :volume averaged strain energy density
ϕ = (1/2)〈�(x)Tσ(x)〉V
5.5 6 6.5 70
1
2
3
4
5
6
7
8
volume averaged strain energy density
coarse scalefine scale
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 15 / 18
Validation : Static
pdfs of QOI in subvolumes :
4 5 6 70
4
8
V11
4 5 6 70
4
8
V12
4 5 6 70
4
8
V13
4 5 6 70
4
8
V21
4 5 6 70
4
8
V22
4 5 6 70
4
8
V23
4 5 6 70
4
8
V31
4 5 6 70
4
8
V32
4 5 6 70
4
8
V33
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 16 / 18
Validation : Ultrasonic wave response
QOI : attenuation coefficient α
I2(xi ) = Io
2 e−2α‖xi−xo‖
0.0 1.5 3.0 6.0
0.0
1.5
3.0
6.0
y (
mm
)
x (mm)
4.5
4.5
0 0.05 0.1 0.15 0.2 0.25−0.5
−0.25
0
0.25
0.5
0.75
1Ricker pulse with fc=10 MHz
Time (µ sec)
Am
plit
ud
e
0 0.5 1 1.5 2 2.5 3−18
−16
−14
−12
−10
−8
−6
−4
−2
0
||xi−xo|| (mm)
log
(I2(
xi)
/ I2o)
computed valueslog(I
2(xi) / I
2o)=(−2α)||xi−xo||
2.8 2.9 30
10
20
30
40
50
60
α (1/mm)
fc=2 (MHz)
1.4 1.60
5
10
15
20
25
30
35
40
45
α (1/mm)
fc=5 (MHz)
0.6 0.7 0.80
5
10
15
20
25
30
35
40
45
50
α (1/mm)
fc=10 (MHz)
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 17 / 18
Summary
1 Developing a statistical tool for characterization and realization of randompolycrystal.
characterization and simulation of geometry with elongated grains.characterization and simulation of Euler angles random field.
2 Adopting a probabilistic model for mesoscale elasticity tensor of polycrystallinematerials.
3 Presenting a validation methodology for prediction of the probabilistic model
Validation in static case.Validation in dynamic case (ultrasonic wave response).
R. Ghanem (USC) SIAM-UQ-2012, Raleigh, NC April 5, 2012 18 / 18