Non-destructive optical depth profiling and real-time evaluation of spectroscopic data

M. Fried, L. Rédei

Research output: Contribution to journalArticle

8 Citations (Scopus)

Abstract

The interpretation of optical measurements is based on a simulation and regression program which minimizes the difference between calculated and measured spectra. A complicated multilayer structure is usually modelled as a system built up of plane-parallel thin films consisting a mixture of two (or more) components. The quality of this interpretation thus depends on the realism of the proposed optical model, on the quality of the reference files for the refractive indices, on the theory of the optical response of mixed layers and on the regression algorithm. For depth profiling a simple model consisting of a few layers can be used to get a first approximation of the thickness of the buried layer and of a possible top layer. The next step is to replace this simple model by a more complex model. Two approaches can be followed for profiles with (a) unknown or (b) known depth variations. Ion-implanted, electrochemically etched silicon and other examples will be presented. Traditionally the evaluation of spectroscopic ellipsometry and reflectometry needs a predetermined multilayer optical model which possesses initial parameters close enough to the targeted ones, because a non-linear gradient descent algorithm uses the initial guess. However, erroneous results may appear when the algorithm falls into the trap of a local minimum. Backpropagation trained feed-forward neural networks proved to be able to give good initial estimations in most cases for avoiding problems. Moreover, a set of neural networks can be trained to perform successive refining of the estimation regions, providing the capability of real-time monitoring of process parameters or fast evaluation of wafer mapping measurements.

Original languageEnglish
Pages (from-to)64-74
Number of pages11
JournalThin Solid Films
Volume364
Issue number1
DOIs
Publication statusPublished - Mar 27 2000

Fingerprint

Depth profiling
optical thickness
evaluation
regression analysis
Optical multilayers
Spectroscopic ellipsometry
Feedforward neural networks
refining
descent
Silicon
Backpropagation
files
optical measurement
laminates
Refining
ellipsometry
Refractive index
Multilayers
traps
wafers

ASJC Scopus subject areas

  • Surfaces, Coatings and Films
  • Condensed Matter Physics
  • Surfaces and Interfaces

Cite this

Non-destructive optical depth profiling and real-time evaluation of spectroscopic data. / Fried, M.; Rédei, L.

In: Thin Solid Films, Vol. 364, No. 1, 27.03.2000, p. 64-74.

Research output: Contribution to journalArticle

@article{78b6742363cd4aa685471d3b24276010,
title = "Non-destructive optical depth profiling and real-time evaluation of spectroscopic data",
abstract = "The interpretation of optical measurements is based on a simulation and regression program which minimizes the difference between calculated and measured spectra. A complicated multilayer structure is usually modelled as a system built up of plane-parallel thin films consisting a mixture of two (or more) components. The quality of this interpretation thus depends on the realism of the proposed optical model, on the quality of the reference files for the refractive indices, on the theory of the optical response of mixed layers and on the regression algorithm. For depth profiling a simple model consisting of a few layers can be used to get a first approximation of the thickness of the buried layer and of a possible top layer. The next step is to replace this simple model by a more complex model. Two approaches can be followed for profiles with (a) unknown or (b) known depth variations. Ion-implanted, electrochemically etched silicon and other examples will be presented. Traditionally the evaluation of spectroscopic ellipsometry and reflectometry needs a predetermined multilayer optical model which possesses initial parameters close enough to the targeted ones, because a non-linear gradient descent algorithm uses the initial guess. However, erroneous results may appear when the algorithm falls into the trap of a local minimum. Backpropagation trained feed-forward neural networks proved to be able to give good initial estimations in most cases for avoiding problems. Moreover, a set of neural networks can be trained to perform successive refining of the estimation regions, providing the capability of real-time monitoring of process parameters or fast evaluation of wafer mapping measurements.",
author = "M. Fried and L. R{\'e}dei",
year = "2000",
month = "3",
day = "27",
doi = "10.1016/S0040-6090(99)00904-9",
language = "English",
volume = "364",
pages = "64--74",
journal = "Thin Solid Films",
issn = "0040-6090",
publisher = "Elsevier",
number = "1",

}

TY - JOUR

T1 - Non-destructive optical depth profiling and real-time evaluation of spectroscopic data

AU - Fried, M.

AU - Rédei, L.

PY - 2000/3/27

Y1 - 2000/3/27

N2 - The interpretation of optical measurements is based on a simulation and regression program which minimizes the difference between calculated and measured spectra. A complicated multilayer structure is usually modelled as a system built up of plane-parallel thin films consisting a mixture of two (or more) components. The quality of this interpretation thus depends on the realism of the proposed optical model, on the quality of the reference files for the refractive indices, on the theory of the optical response of mixed layers and on the regression algorithm. For depth profiling a simple model consisting of a few layers can be used to get a first approximation of the thickness of the buried layer and of a possible top layer. The next step is to replace this simple model by a more complex model. Two approaches can be followed for profiles with (a) unknown or (b) known depth variations. Ion-implanted, electrochemically etched silicon and other examples will be presented. Traditionally the evaluation of spectroscopic ellipsometry and reflectometry needs a predetermined multilayer optical model which possesses initial parameters close enough to the targeted ones, because a non-linear gradient descent algorithm uses the initial guess. However, erroneous results may appear when the algorithm falls into the trap of a local minimum. Backpropagation trained feed-forward neural networks proved to be able to give good initial estimations in most cases for avoiding problems. Moreover, a set of neural networks can be trained to perform successive refining of the estimation regions, providing the capability of real-time monitoring of process parameters or fast evaluation of wafer mapping measurements.

AB - The interpretation of optical measurements is based on a simulation and regression program which minimizes the difference between calculated and measured spectra. A complicated multilayer structure is usually modelled as a system built up of plane-parallel thin films consisting a mixture of two (or more) components. The quality of this interpretation thus depends on the realism of the proposed optical model, on the quality of the reference files for the refractive indices, on the theory of the optical response of mixed layers and on the regression algorithm. For depth profiling a simple model consisting of a few layers can be used to get a first approximation of the thickness of the buried layer and of a possible top layer. The next step is to replace this simple model by a more complex model. Two approaches can be followed for profiles with (a) unknown or (b) known depth variations. Ion-implanted, electrochemically etched silicon and other examples will be presented. Traditionally the evaluation of spectroscopic ellipsometry and reflectometry needs a predetermined multilayer optical model which possesses initial parameters close enough to the targeted ones, because a non-linear gradient descent algorithm uses the initial guess. However, erroneous results may appear when the algorithm falls into the trap of a local minimum. Backpropagation trained feed-forward neural networks proved to be able to give good initial estimations in most cases for avoiding problems. Moreover, a set of neural networks can be trained to perform successive refining of the estimation regions, providing the capability of real-time monitoring of process parameters or fast evaluation of wafer mapping measurements.

UR - http://www.scopus.com/inward/record.url?scp=17044452327&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=17044452327&partnerID=8YFLogxK

U2 - 10.1016/S0040-6090(99)00904-9

DO - 10.1016/S0040-6090(99)00904-9

M3 - Article

VL - 364

SP - 64

EP - 74

JO - Thin Solid Films

JF - Thin Solid Films

SN - 0040-6090

IS - 1

ER -