Introduction

Properties

Large aspect ratio(>1000).

Atomically Sharp tips.

High temperature and chemical stability.

High electrical and thermal conductivity.

Can be either Metal or semiconductor.

Problems in the Field

Experimental data doesn’t match the theoretical

expectations due to quality issues.

Comprehensive theoretical models are still an

approximation, which doesn’t involve the

distinctiveness of every sample.

Methods for deterministic (controlled) synthesis

and assembly of carbon nanostructures for novel

devices and materials have not been developed.

“Nanotube suppliers accused of selling shoddy

goods” by Jim Giles “ Researchers who buy products

such as carbon nanotubes are frequently being sold

defective materials, according to a survey of

nanotechnology companies. ” [news@nature published

online: 10 December 2004 | doi:10.1038/news041206-15.]

Raman SpectroscopyNondestructive, lots of information content

(wavelength, polarization, batch average over micron

scale). Research has proven that some geometrical

information are embedded in Raman Spectrum [2].

Image Processing

2D-FFT (Fast Fourier Transform) analysis.

2D Stereological Relations:

Relative Alignment measurements:

Measured based on the variance of the relative distances di, between even

segments, si, of any two nanotubes.

Where N is the total number of segments, and M is the number of two

nanotube combinations.

Relative Curvature measurements:

The curvature, k, at considered segments is defined

as the rate of rotation, , of the nanotube tangent

as the contact point moves along the segment length, s:

Artificial Neural NetworksBackpropagation training methodology is used in

training the type of neural network [3].

Quality Analysis MethodologyRaman spectroscopy and image processing analysis are combined to represent

the sample geometrical characteristics. This empirical model will assist the

quality control process of CNT production, enabling manufacturers to achieve

desired properties by monitoring the difference between both desired CNT

analysis and the CNT under nucleation. The flow chart in fig.4 shows the main

blocks of the implemented procedure.

Technical ProcedureThe procedures for analyzing nanotubes are:

1)Filtering and thresholding the SEM image, and enhancing the signal-to-noise

ratio in Raman spectra using Wavelets de-nosing techniques.

2)Extracting the morphology of the nanostructure using image processing

techniques of Scanning Electron Microscope (SEM) images, stereological

relations and curvature and alignment measurements.

3)Raman spectroscopy investigation of the characterized CNT specimens is

performed using multiple excitation wavelengths.

4)The relationship between the experimental spectroscopic analysis and the

morphology of the structure will be modeled using Artificial Neural Networks

(ANN).

Experiment ProcedureSEM Imaging

Raman Spectroscopy

CharacterizationImage Processing

Fig1. CNTs structures

FFT Analysis

FFT was applied to both real images and ideal images, to understand the meaning

of the results based on CNTs alignment, curvature, orientation, and thickness

properties.

Stereological Analysis

Raman Spectrum

Filtering using Wavelet data processing [4]

Peaks info.

Empirical ModelingArtificial neural network model for CNTs quality classification

ConclusionModeling using Raman Spectroscopy analyses and

SEM image processing will link CNT properties with

growth conditions, which could be used later to

manipulate the nucleation factors such that the CNT

desired properties in large quantities will be

achieved.

Neural networks classifier is very helpful to

categorize the images based on alignment and purity

characteristics. Empirical modeling is a powerful tool

to predict, explain, and model materials behavior,

especially when the analytical modeling is not

achievable.

Fig3. Artificial Neural Network Architecture

1

1

1

,

1 k

i

k

j

s

j

k

ji

k

ibawn

k

)( 111 k

i

kk

infa

1))*2exp(1(

2)(

nnTansig

k

ji

k

ji

w

Vw

,

,

Quality Classification via Raman Identification of Carbon Nanotube Bundles Using Artificial Neural Networks

M.Al-Khedher¹, C. Pezeshki¹, J. McHale², F. Knorr²¹ School of Mechanical and Materials Engineering, ² Department of Chemistry

Washington State University, Pullman, WA 99164-2920, USA

Fig7. Auto-Threshold (a) raw image (b) correlation vs. Threshold (c) Thresholded image at max

correlation.

Fig.8 Different FFT patterns of created ideal images (a-c) and SEM images (d-e).

Fig10. (a) Raw portion of Raman signal (b) DWT and IDWT data processing (c) filtered Raman signal.

Fig12. Model outputs for (a) sample1 and (b) sample2 alignment estimations, (c) sample1 and (d) sample2

curvature estimations. Insets are the SEM images of the analyzed samples.

AbstractThe exceptional properties of Carbon Nanotubes

(CNTs) have attracted much attention in recent years

for their small dimensions, high electrical and thermal

conductivities and unique morphologies [1].

It is important to control the structure of these CNTs if

we want to control the properties they exhibit for

mass-production purposes. The methodology is

basically correlating the morphology of the structure,

which is extracted using image processing techniques

of Scanning Electron Microscope (SEM) images, and

Raman Spectroscopy analysis with the mechanical

properties of the structure using Artificial Neural

Network modeling technique. The model will play a

significant role in analyzing and predicting the CNTs’

properties, which will eventually help in designing the

desired CNT structure.

Fig5. SEM images of the turf (a) corner view (b) detail (c) SEM of different samples

• Area Fraction:

• Perimeter length per unit area:

• Line intercept count:

w w

CNTA

A

AA

probe

tionserL

L

PP secint

w w

boundaryCNT

AA

LL

_

Fig2. Raman Spectroscopy of SWNT & MWNT

Raman Spectroscopy

of CNTs

Quantitative Analysis

of Raman Spectrum

SEM Images of CNTs

turfs (bulk material)

Image Analysis for

Numerical Representation

of Images

Artificial

Neural Network

Model for CNTs

Quality

Classification

Artificial

Neural Network

for

Raman Spectrum

Identification

Fig.4 Scope of the project.

References

[1] P Harris, Carbon Nanotubes and Related Structures,

Cambridge University Press , 1999.

[2] Rao, A. M., Richter, E., Bandow, S., Chase, B.,

Eklund, P. C., Williams, K. A., Fang, S., Subbaswamy,

K. R., Menon, M., Thess, A., Smalley, R. E.,

Dresselhaus, G., and Dresselhaus, M. S. Diameter-

selective raman scattering from vibrational modes in

carbon nanotubes, Science 275, pages 187-191 1997.

[3] K K Tho, S Swaddiwudhipong, Z S Liu and J Hua,

Artificial neural network model for material

characterization by indentation, Modelling and

Simulation in Materials Science and Engineering Vol.

12 Issue 5 Article 019 2004-09-01.

[4] C Camerlingo, F Zenone, G M Gaeta, R Riccio and

M Lepore,C Camerlingo et al Wavelet data processing

of micro-Raman spectra of biological samples, Meas.

Sci. Technol. 17 298-303 2006.

Fig6. Raman Spectra of two different CNT samples taken at two excitation

wavelengths: 488.0 and 530.9 nm. The shown range is: 600cm-1 -3400cm-1.

M

1j

N

1i

2i

M

1j

i ))dd(N

1(

M

1))d(var(

M

1A

ds

dk

IDWTDWT

2500 2600 2700 2800 2900 3000 310020

30

40

50

60

70

80

90

100

Relative wavenumber (cm-1)

Rela

tive I

nte

nsity(c

ounts

)

2500 2650 2800 2950 3100

D8

D7

D6

D5

D4

D3

D2

D1

(b)

(c)(a)

Fig11. Lorentzian fitting of 2D band at 2718cm-1.

0 50 100 150 200 250 3000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Threshold

corr

ela

tion

image after thresholding at the peak correlation

500 1000 1500 2000 2500

200

400

600

800

1000

1200

1400

1600

1800

(a) (b) (c)

Discussion• Understanding the correlation between CNT growth

aspects and its properties is the key factor to develop

quality control methods for controlled synthesis and

assembly of carbon nanotubes.

• The system is built to model and estimate CNTs

quality based on Raman Spectroscopy, morphology

information extracted from SEM images using in-

house image processing analysis.

• FFT statistics were a good indicator for CNTs

alignment, curvature, orientation, and thickness

properties.

• Wavelet de-noising analysis produced a smoothed

Raman signal without loosing any peaks information

embedded in the spectrum.

• Artificial neural network model for CNTs alignment

classification using image processing results shows an

accurate categorizing technique for CNTs with an

acceptable range of approximation.

• ANN was able to model the Raman data and extract

the influence of the given inputs. The effect of the

excitation wavelength and CNT morphology was

intensely studied.

(a) (b) (c)

(a1)

(a2)

(b1)

(b2)

(c1)

(c2)

(d1)

(d2)

Fig.9 (a) SEM image samples of analyzed specimen for the measurements of: (b) AA (C) LA (d) PL( ).

Quality evaluation using Raman Spectra

(a) (b)Fig.13 Effect of excitation wavelength on (a) D, (b) 2D peak positions.

Fig.14 ANN results for the effect of (a) LA (b) PL on the HWHM of G line.(a) (b)

(e)(d)

image after thresholding at the peak correlation

image after thresholding at the peak correlation

Fig.15 Neural network results of the D line (a) Upshift in D band as a

function of PL , (b) Upshift in 2D band as a function of LA

(a) (b)