1.

Introduction
Surface topography impacts functionality of the work piece [13], consequently surface
metrology is becoming more and more important. Optical profilers are used to measure surface
topography in micro and nano-scale. The majorly used optical profilers were scanning white
light interferometers (SWLI), whose vertical resolution can be sub-nm [4], and the scanning
confocal microscope which offers the advantage of a larger measurable slope range, reportedly
as large as

[5]. The measurement of surface topography involves XY mapping of height

data.
Stage drift during the measurement causes errors in the height and x y mapping of the height
data. Stage drift has long been a limiting factor in tracking micro features of surfaces during a
scanning process [6]. For sequential frame data acquisition, drift will lead to a mismatch between
subsequent measurements. Thus, drift rate considerations are important and often need to be
reduced through environmental, procedural, and instrument modifications. To characterize the
instrument performance and to improve the instrument, instrument drift should be characterized
and repeated often and over a range of time scales, as each instrument in its own environment
has its own unique drift characteristics. In this paper, we discuss the combination of a spherical
artifact and a robust ball fitting algorithm to estimate drift rates.

2. Existing Methodologies for drift calibration

Several methods can be used to address instrument drift, and they can be roughly grouped
into two categories: drift calibration with target tracking and drift calibration with active
feedback systems.

2.1 Drift Calibration with target tracking:

There are several approaches in the target tracking category [79]. The simplest approach
is tracking a fiduciary marker which is marked on the top of sampling. Resolution on the
order of 1 nm in the horizontal XY plane and 5 nm in Z direction has been demonstrated in
certain limits [10]. The preparation of the marker and the setup are often complex, and this
method is not always practical for surface metrology needs in precision engineering,
especially when stage drift changes from day to day.

In addition, tracking the targets position is not an easy task as measurement noise can
easily affect data processing and the identification of the targets location.

2.2 Drift calibration with active feedback systems.

Active feedback systems are another way to stabilize the stage drift. In the setup, a
marker was used to track displacement and a piezoelectric stage was used to correct drift. By
using this method a system can be made that has stability of 0.1nm in three dimensions for 1
second [11]. Setups are complicated in this category and usually costly.
The existing methodologies are too costly and not so convenient. There will be measurement
uncertainties as the readings are taken in small part of field of view (tracks only target). In many
applications, a convenient and robust method for simply calibrating drift rates is sufficient.

3. Proposed Methodology for drift calibration

In this paper, we discuss the combination of a spherical artifact and a robust ball fitting
algorithm to estimate drift rates. Instead of tracking a visible spot in the field of view (FOV)
directly, we take subsequent height profile measurements of a cap on a spherical artifact whose
diameter is larger than the field of view. We use a robust fitting algorithm to find the center of
the best-fit sphere to the data and the center location provides the information on the drift
properties, such as drift direction and drift rate.
Drift causes the ball to move over the field of view so tracking the position of the ball center
provides the drift information. For short periods of time, drift is usually linear [12]. Let x and y
represent the coordinates in the plane of the part and z represent the orthogonal height axis. The
drift model will be

Where

and

( )

(1)

( )

(2)

are the center coordinates of the best-fit sphere and

and

are the

coordinates at the start of the drift test (approximately the middle of the field of view). The linear
drift rate can also be evaluated as

(3)

The center coordinate tracked along a line as

2

(4)
The drift rates are determined by taking a series of measurements of the surface of the
spherical artifact, fitting each measurement to a sphere to determine the center coordinates, and
plotting the center coordinates as a function of time. A linear fit to the data is then used to extract
the slopes x and y for Eqs. (1) and (2).
Since the method relies on determining the center position of the ball from measurements of
a spherical cap on the ball, several factors must be considered such as the algorithm used to
determine the best-fit sphere to the spherical cap data, the realistic geometry of the ball, the size
of the ball, and random noise in each measurement.

3.1 Ball size

Choosing the best diameter for the spherical artifact is important. If the diameter is
too large, a spherical fit becomes ill defined as the height profile becomes too flat over the
field of view. If the diameter is too small, steep slopes at the edge of the field of view can
lead to noisy data and/or data dropout. Best calibration occurs when the ball is as small as
possible with minimal data dropout.

Fig.1. Relative size of scales for selection of ball diameter.

The technique uses height profile data from a cap on the surface of a ball to estimate
the center of the best-fit sphere for the entire ball. Given geometric errors on a ball and noise
in each measurement, the best estimate of the balls center occurs when the size of the cap is
large. All profilometers have a maximum measurable slope. The slope is large at the edge of
3

the field of view, so if the cap is too large, the instrument will not be able to measure these
regions as shown Fig. 2. Thus, the smallest diameter ball should be used that allows a surface
measurement over the entire FOV. Since measurable slopes are usually quite limited, this
typically leads to a ball diameter choice that is larger than the field of view, as shown Fig. 1.

Fig.2. Schematic of ball profile when balls of various diameters are used

3.2 Best Fit Algorithm

Least-square fitting to a cap on a sphere to determine the best fit center must be done
with care and a robust best-fit algorithm is needed. We use a method proposed by Forbs [13]
(Fig. 2). The basic idea for the spherical cap fit is to constrain the number of parameters of
the best-fit sphere model [14,15].
In essence, a general equation of a spherical cap in Cartesian coordinates can be written
as
(

(5)

where (x, y, z) is the point on a sphere, and A, B, C, D, E are the coefficients of an

equation for a sphere. A plane can be fit to the height profile of the cap and rotated so the
normal to the best-fit plane points along the positive Z axis, this ensures D = 1 in the
equation. Then a least square fit is carried out to determine the best fit radius and coordinates
of the sphere center.
4

3.3 Geometrical Errors on the ball

Geometric errors on the ball will impact the center coordinates of the best-fit sphere
when fitting over a finite area. The goal of the drift measurement is to track the ball center
coordinate change over time, so an error in the best-fit center coordinates is only a problem if
the error changes over time. This is a concern only if the ball geometry over the field of view
changes significantly over time. This is best explored through simulation and is described in
the simulation section

3.4 Random Noise

Random noise in each measurement will impact the drift rate measurement and the
ball size choice. There are many causes of random noise, e.g. air turbulence and noise from
the scanner, and each has a unique spatial frequency characteristic. Noise sources with low
spatial frequency content are the most important to consider as these will have the largest
impact on the center coordinates of a best-fit sphere to the height profile data. The magnitude
and spatial character of the random noise should be evaluated to understand this impact on
the drift rate measurement. This can be done with a repeatability test for the instrument
where a sequence of ball profile measurements are taken without changing the ball position
as quickly as possible (to minimize drift between measurements). The random noise can be
assessed by calculating the difference between subsequent measurements and repeating this
for many measurement pairs to obtain a statistical sample of the random noise characteristics.
Each measurement

be written as
(6)

()

where

is the profile of a perfect ball,

is the ball error on the local patch,

is the systematic error from the instrument itself, and

()

is the

random noise term. The first three terms on the right in Eq. (6) are constants.
Pair wise subtraction of adjacent measurements indicates the random noise contributions
only
(

()

Since we are most interested in low-order spatial frequency noise, the set of low-order
Zernike polynomials are useful for the analysis. The first 36 lowest order Zernike
5

polynomials [16] can be fit to each noise difference map, and then the statistical distribution
of the coefficients extracted. The standard deviation for each coefficient can be used in a
simulation to randomly generate noise to then add to each simulated measurement (assuming
a Gaussian probability distribution).

Fig.3. Decomposition of a noise difference map .

We carried out this noise evaluation for a confocal microscopy with a 50X objective.
Twenty measurements were taken and 10 noise difference maps were generated by
subtracting every pair of subsequent measurements. We then decomposed each difference
map into low and high-order spatial frequency components by fitting with the first 36
Zernike polynomials. We found that the random noise consists of a low frequency and a high
frequency component, as shown in Fig. 3. The ring structure in the high frequency noise map
(Fig. 3. Bottom right) is likely noise from the scanner during the data acquisition process.
The low frequency noise (Fig. 3. Top right) is likely from environmental effects such as air
turbulence. From the spatial signature ofthe two noise contributions, it is clear that the lowfrequency noise will have the largest impact on the spherical best-fit analysis

4. Simulation
Simulated measurements were carried out in Matlab to test the validity and limits of the
proposed method. The simulation starts by defining a realistic ball geometry using a spherical
harmonic basis set (Fig. 1) and then positioning the ball in a desired location in a defined field of
view. The sag (height) values over the field of view are then taken as the simulated profilometer
measurement. Drift can be applied to the ball location in the field of view and random noise can
be added to each measurement. Each measurement is then with the algorithm which finds the
best-fit sphere to each simulated cap measurement. The best-fit sphere center coordinates are
then plotted versus time to extract the drift rates

and

(Eqs.(1) and (2)).

In the simulation, the drift test was set as 100 min with one measurement per minute, the
trace of the balls center coordinates as a function of time therefore determine the drift property.
The uncertainty of the drift rate is determined by the standard deviation of the drift values The
simulation shows that, with R = 0.595 mm using the 50X objective whose field of view is 0.256
mm X 0.256 mm, the uncertainty of the drift rate measurement

in both the X and Y directions

is on the order of 0.5 nm/min, as shown in Fig. 4.

Fig.4. Simulation result of drift in x and y direction

We explored this in simulation by fixing the field of view (corresponding to the 50X
objective condition), fixing the magnitude of the random noise, and varying the radius of the
7

simulated ball, drift test was set as 100 min with one measurement per minute. The same
procedure discussed above was followed to estimate a drift rate uncertainty for each ball radius
considered. The results are summarized in Fig. 5 were

for both X and Y.

Fig.5. Drift uncertainty of confocal microscope as a function of ball size

We can observe from Fig.5. As the ball radius is reduced, data is dropped out of the
simulated measurements when the slope becomes larger than the reported measurement limit for
the confocal microscope. Thus the measurable slope limitation becomes important in the small
ball limit and the uncertainty will again rise. For the conditions presented here, this effect is
smaller and becomes significant only for very small balls.
In summary, the simulation shows that a reasonable ball size with reasonable quality can
be used to measure drift rates with reasonable measurement uncertainties. With a ball with radius
R = 0.595 mm a drift test over 100 min with one measurement per minute results in drift rate
uncertainties of

0.8 nm/min.

5. Experimental results
In the experiment, a commercial confocal microscope (Olympus LEXT OLS4000)
located in a moderately temperature-controlled clean room was used to demonstrate the
procedure. A grade 3 steel ball with R = 0.595 mm was chosen for the measurement. A simple
8

fixture was used to allow the ball to be stabilized, consisting of a brass sheet with a center
indentation sandwiched between two large washers. The ball rests at the indention of the fixture
that is placed on the microscope stage, and the fixture is allowed to acclimate before the
measurement. It is approximately centered in the field of view at the start of the drift test. Using
the 50X objective, 14sequential measurements were taken and each measurement took
approximately 1 min to acquire. The time interval between measurements and the duration of the
drift test will vary and depend on the application. Each measurement is processed as described
above to extract the center coordinates of the best fit sphere. The center coordinates are plotted
versus time, as shown in Fig. 6, and fit to a line to estimate the drift rate. The uncertainty of the
drift rate was determined by the standard deviation of the fitted slope values.

Fig.6. Experimental results of confocal microscope drift detection

In our test, the X drift rate was 75 nm/min with uncertainty

rate was 54 nm/min with

and the Y drift

The drift is approximately linear over this time

scale forthis instrument. The relatively larger uncertainty in the drift rate in the experiment
compared to simulation is due at least in part to the shorter duration of the drift test in the
experiment (14 mincompared to 100 min for the simulation).

6. Conclusion:
Drift is present in all instruments, and contributes to measurement uncertainty at some level.
For measurements where time-lapse information is important, the uncertainty caused by drift can
be large. We proposed and demonstrated a non-traditional way to characterize the drift in an
instrument that is based on tracking the center of the best fit sphere to measurements of the
surface of a ball artifact. Simulation shows that choosing a ball with a suitable ratio of field of
view to ball size is important and that the method can be quite effective in terms of low
uncertainty. As important, the method is simple and economical to implement. Due to the
arbitrary vertical offset of the scanner from one scan to another, this method is not suitable for
determining the vertical drift rate (in the scan direction).

References:
[1] Whitehouse DJ. Surface metrology. Measurement Science & Technology 1997;8(9):95572.
[2] Bruzzone AAG, Costa HL, Lonardo PM, Lucca DA. Advances in engineered surfaces for
functional performance. CIRP Annals Manufacturing Technology 2008;57(2):75069.
[3] Jiang XJ, Whitehouse DJ. Technological shifts in surface metrology. CIRP Annals
Manufacturing Technology 2012;61(2):81536.
[4] Deck L, de Groot P. High-speed noncontact profiler based on scanning whitelight
interferometry. Applied Optics 1994;33(31):73348.
[5] Olympus LEXT OLS4000 laser scanning confocal microscopy manual.
[6] Adler J, Pagakis SN. Reducing image distortions due to temperature related microscope stage
drift. Journal of Microscopy 2003;210(2): 1317.
[7] Lee Sang Hak, Baday Murat, Tjioe Marco, Simonson Paul D, Zhang Ruobing, Cai En, et al.
Using fixed fiduciary markers for stage drift correction. Optics Express 2012;20(11):1217783.
[8] Gelles J, Schnapp BJ, Sheetz MP. Tracking kinesin-driven movements with nanometre-scale
precision. Nature 1988;331(6155):4503.
[9] Dangaria JH, Yang S, Butler PJ. Improved nanometer-scale particle tracking in optical
microscopy using microfabricated fiduciary posts. Biotechniques 2007;42(4):437.

10

[10] Betzig Eric, Patterson George H, Sougrat Rachid, Lindwasse OWolf, Olenych Scott,
Bonifacino Juan S, et al. Imaging intracellular fluorescent proteins at nanometer resolution.
Science 2006;313(5793):16425.
[11] Carter AR, King GM, Ulrich TA, Halsey W, Alchenberger D, Perkins TT. Stabilization of
an optical microscope to 0.1 nm in three dimensions. Applied Optics 2007;46(3):4217.
[12] Olympus LEXT OLS4000. Charlotte: Physics and Optical Science Department, UNC.
[13] Forbs AB. Robust circle and sphere fitting by least squares; 1989
[14] Zhou Y, Ghim Y-S, Davies A. Self-calibration for slope-dependent errors in optical
profilometry by using the random ball test. Proceedings of the SPIE 2012. Article ID 84930H.
[15]

Samara

AM.

Enhanced

dynamic

range

fringe

projection

for

micro-structure

characterization; 2005.
[16] Wyant JC, Creath K. Basic wavefront aberration theory for optical metrology. Applied
Optics and Optical Engineering 1992;11(s 29):2.

11

Table of Contents
Abstract ............................................................................................................................................ i
1.

Introduction ............................................................................................................................. 1

2.

Existing Methodologies for drift calibration ........................................................................... 1

3.

2.1

Drift Calibration with target tracking: .............................................................................. 1

2.2

Drift calibration with active feedback systems. ............................................................... 2

Proposed Methodology for drift calibration ............................................................................ 2

3.1

Ball size ............................................................................................................................ 3

3.2

Best Fit Algorithm ............................................................................................................ 4

3.3

Geometrical Errors on the ball ......................................................................................... 5

3.4

Random Noise .................................................................................................................. 5

4.

Simulation................................................................................................................................ 7

5.

Experimental results ................................................................................................................ 8

6.

Conclusion: ............................................................................................................................ 10

References: .................................................................................................................................... 10

12