CIRP Annals - Manufacturing Technology 68 (2019) 531–534

Contents lists available at ScienceDirect

CIRP Annals - Manufacturing Technology

jou rnal homep age : ht t p: // ees .e lse vi er . com /ci r p/ def a ult . asp

Uncertainty of particle size measurements using dynamic image

analysis$
Justin G. Whiting a, Vipin N. Tondare a,b, John Henry J. Scott c, Thien Q. Phan a,
M. Alkan Donmez (2)a,*
a
Engineering Laboratory, National Institute of Standards and Technology (NIST), Gaithersburg, Maryland 20899, USA
b
Theiss Research, La Jolla, California 92037, USA
c
Material Measurement Laboratory, NIST, Gaithersburg, Maryland 20899, USA

A R T I C L E I N F O A B S T R A C T

Metal powder particle size distribution (PSD) is a critical factor affecting powder layer density and
Keywords: uniformity in additive manufacturing processes. Among various existing measurement methods,
Additive manufacturing dynamic image analysis (DIA) instruments are very appealing for measuring PSD. However, the ‘black box’
Uncertainty nature and complex measurement process inherent to DIA make quantification of uncertainty
Particle size distribution challenging. A method to establish DIA-based measurement uncertainty based on calibrated powder
samples via a scanning electron microscope is described. Uncertainty analysis was performed taking into
account uncertainties associated with the calibration of the sample as well as non-similarities of the
calibrated sample and the measured sample.
© 2019 Published by Elsevier Ltd on behalf of CIRP.

1. Introduction left by larger ones, they are more susceptible to Van der Waals
forces and create a more cohesive powder which can lead to poorly
Metal-based additive manufacturing (AM) has great industrial spread layers [3].
potential in a wide range of industries including aerospace, While a variety of industrially common particle size measure-
automotive, and biomedicine. However, to realize this potential, it ment methods exist such as sieve analysis, laser light scattering,
must demonstrate the capability to consistently fabricate high static image analysis, and dynamic image analysis (DIA), there is a
quality industrial parts. Metal-based AM is sensitive to metal lack of information detailing the uncertainty of these methods,
powder quality [1,2]. Poor powder quality can produce defects in particularly when they are applied to sizing gas-atomized metal
the end part including pores, cracks, inclusions, residual stresses, powders like those used in AM. Furthermore, when common
and suboptimal surface roughness. Particle size and particle size sizing methods were compared, they produced significantly
distribution (PSD) of the metal powder used in additive different results [4]. The DIA method is attractive for its relative
manufacturing are among the most influential factors. For low cost, measurement speed, and ability to provide morpholog-
example, proper flowability of metal powder during the AM ical information. In its most basic form, the DIA measurement
process depends on particle size, shape, and PSD among various principle, involves establishing contours of moving particles by
other environmental conditions (e.g., moisture) and AM machine capturing their silhouettes (shadows) using a high-speed camera
properties (e.g., dynamic characteristic of recoating mechanism). following the computational theory developed in the 1990s
In general, smooth, spherical particles tend to flow more easily and [5]. Using either a liquid or gaseous dispersion, the DIA
therefore produce uniform, high-density layers. Additionally, instrument conveys a stream of sample particles past a light
maximum packing density of particles in a formed layer is source, and images are taken of the projected shadows of the
achieved with a specific PSD that includes both coarse and fine particles. The images undergo processing and eventually are
particles. While finer particles increase density by filling the gaps converted to grey-scale images so the two-dimensional (2D)
morphology and various size metrics of the particle population
can be extracted.
$
Although the basic operating principle is well established and
Official contribution of the National Institute of Standards and Technology, not
standardized [5,6], as shown in Fig. 1, the particular means
subject to copyright in the United States.
* Corresponding author. employed can be complex and can vary depending on the
E-mail address: [email protected] (A. Donmez). instrument manufacturer. For example, since the powder is in a

https://doi.org/10.1016/j.cirp.2019.04.075
0007-8506/© 2019 Published by Elsevier Ltd on behalf of CIRP.
532 J.G. Whiting et al. / CIRP Annals - Manufacturing Technology 68 (2019) 531–534

DIA-based ECAD measurements at these three levels were

calculated from the derived uncertainties in corresponding
frequencies by Monte Carlo simulations.
Multiple powder samples with nominally identical PSDs and
morphologies representative of the bulk metal AM powder were
prepared using a rotary riffler [8].

Fig. 1. (a) Flow diagram for typical DIA method: 1. dispersed particles 2. device for
control of particle motion 3. measurement volume 4. light source 5. optical system
6. depth of field 7. image capture device 8. image analyzer and 9. PSD display. (b) DIA
Schematic.: 1. light source 2. camera and 3. measurement volume. (adapted from
Ref. [6]).

Fig. 2. (a) Illustration of SEM sample preparation apparatus (not to scale) and, (b) a
typical SEM micrograph of the stainless-steel metal particles. Horizontal full width
dynamic state, both the illumination and the exposure duration of the SEM image is 1.06 mm.
will impact what the image sensor transduces into a signal. This
is further complicated by the fact that the powder is entrained by
a gas or liquid, and the drag force will produce a gradient of 2.1. SEM sample preparation
particle speeds dependent on the size and shape of the particles
in the flow. A prerequisite for a proper PSD analysis by an SEM is a well-
While this potential source of measurement error is, to some dispersed powder sample. Each dispersion should have a sufficient
extent, alleviated by a compensating algorithm, the proprietary number of particles to be statistically significant while minimizing
nature of commercially available DIA instruments prevents the contact between particles. The SEM sample preparation apparatus
user from knowing the means or extent of the compensation. This used for the metal powder in this work is illustrated in Fig. 2(a). It
velocity gradient is just one of the sources of uncertainty. includes a carbon tape adhered on a metal substrate covered with a
The sources of uncertainty in a DIA-based measurement [6] glass beaker, a curved-tip glass eye-dropper holding the powder
include optical distortion error (pincushion, barrel, etc.), depth of sample, and a manually operated air bulb to blow the powder
field of camera focus/defocus algorithm, motion blur (particle sample through the eye-dropper onto the carbon tape. All
velocity and integration time), particle velocity as a function of components were sealed with rubber joint compound.
size, contaminations, overlapping of particles, incomplete imaging The manually operated air bulb fitted with a check valve to
of particles (particles at edge of image), particle orientation, and prevent backflow was forcefully squeezed, and then the pressure
insufficient number of particles measured. Furthermore, during on the air bulb was released slowly (over about 20 s) while using
image analysis the sources of uncertainty can be from pixel size one hand to stabilize the beaker. This air bulb dispersion procedure
calibration error, background correction (for non-uniformity of was repeated three to five times to ensure no powder was left in
illumination), particle edge determination (thresholding, edge the curved tip and all powder was dispersed onto the SEM
detection), and particle size determination along different substrate. Following this procedure, two SEM samples of dimen-
orientations. Consequently, there is an urgent need to carry out sions (50 mm
50 mm) were prepared for SEM imaging. Fig. 2(b) is
complete and rigorous uncertainty analysis of the particle size and an example of a typical SEM micrograph showing uniformly
PSD measurements using the DIA method. However, due to the dispersed metal particles.
proprietary nature of image acquisition, processing, and analysis
used by commercial DIA instruments, quantification of uncertainty 2.2. SEM data acquisition
is challenging. In such cases, calibrated artifacts are used to assess
measurement uncertainty [7]. But uncertainty associated with the SEM images were acquired in an automated mode, without
differences between the calibrated artifacts (e.g., spherical glass overlap, and using the SEM’s annular backscattered electron
beads) and metal powders used in AM (with complex shapes and detector. Optimum imaging parameters were determined by trial
material optical properties) makes this approach impractical. and error as a beam landing energy of 15 keV, an electron beam
Therefore, we used a variation of this approach to establish DIA- current of 8.9 nA, and a dwell time of 2 ms per pixel. To increase the
based PSD measurement uncertainty based on AM metal powder depth of field and minimize defocus problems, the working
samples calibrated via a scanning electron microscope (SEM). distance was set at 9.8 mm. The horizontal full width, set at
1.06 mm and resulting in pixel size of 0.518 mm, was considered
2. Materials and methods sufficient for imaging powder particles with sizes of tens of
micrometers. The choice of dwell time and pixel size was also a
In this study, a commercially available 17-4 stainless steel trade-off between the number of SEM images and the time
argon-atomized powder was used. A commercial DIA system and required to acquire them.
an SEM were used for PSD measurements with the measurand of Eqs. (1–4) from the reference [9] were used to determine the
equivalent circular area diameter (ECAD). The typical PSD minimum number of particles necessary for achieving at least a
specification (reported in AM powder certification sheets) is given 95% probability level (P) with no more than 3% relative error.
by the cumulative frequency distribution as a function of the size of
2
the particles. However, PSDs are generally reported as a volume n ¼ vd ; ð1Þ
fraction, rather than as a number fraction, of the particles. To
generate such PSD functions, ECAD values for particles were first v ¼ p2 a2 s2 ð2c2 s2 þ 1Þ; ð2Þ
converted to volume assuming a sphere with the given diameter.
Particle volumes were binned with respect to their ECAD values. fðjpjÞ ¼ ð1  PÞ=2; ð3Þ
Volume fraction-based PSDs were then calculated as the sum of the
volumes for each size bin divided by the sum of all particle c ¼ b þ a=2; ð4Þ
volumes. The cumulative frequency distribution function generat-
ed from these bins is used to determine sizes corresponding to the where n is the minimum number of particles and d is relative error.
10th, 50th, and 90th percentiles, which are labelled as D10, D50, In Eqs. 2–4 p, an intermediate parameter, is derived from a
and D90, respectively. Therefore, the final uncertainty estimates in cumulative distribution function, f; of the standard normal
 J.G. Whiting et al. / CIRP Annals - Manufacturing Technology 68 (2019) 531–534 533

distribution. a and b are constants defined from the desired metric in bin frequencies (% by volume) for each parameter were
(e.g., mass median or Sauter diameter) and s is the standard considered as a full range of variation. Standard uncertainties
deviation of the population, which can be derived from the 84th were calculated from these range values for each size and each
and 50th percentile assuming log-normal distribution by number parameter assuming variations have uniform distribution (Fig. 3).
(i.e., s ¼ x84;0 =x50;0 ). In addition to the above-mentioned sources of uncertainty, the
We found that 59 866 particles were sufficient for achieving a most important image processing uncertainty to address is random
95% probability level and 3% relative error in measuring the mass variation in pixel segmentation, i.e., determination of the location in
median diameter of this argon atomized metal powder. In this the image of the boundary between the particle foreground and the
study, two SEM samples, each 50 mm
50 mm, were imaged, and substrate background. The SEM image acquisition strategy was
in total 4575 SEM images were acquired; this acquisition strategy designed specifically to minimize this uncertainty by selecting the
yielded SEM images of more than 69 000 particles. backscattered electron detector operated in composition mode (not
topographic mode) instead of the more commonly used Everhart–
2.3. SEM image processing Thornley detector, thus relying on a contrast mechanism based on the
large difference in atomic number of the metal particles compared to
Each of the acquired SEM images underwent the same the low-Z carbon substrate. Further, the gain and offset of the detector
processing. First, the image was converted to an 8-bit grey-scale amplifier were adjusted to deliberately lower the dynamic range in
format which lowers the computational cost without sacrificing the image, thus acquiring images in hardware close to the ideal
image information content. A Kuwahara filter [10] with window binarized data needed for input into the particle shape and size
size of 5 pixels was then used to smooth the intra-particle features analysis software. While this approach sacrifices fine spatial detail in
while maintaining sharp contrast at the particle’s edges. The image the image of the particle surface (not needed in this study), it has the
was then binarized with a lower threshold value of 13, which was advantage of greatly reducing the sensitivity of the segmentation
empirically found to maximize particle detection while not result to random variations and the choice of particular parameters
including the SEM tape’s surface features. An open-source and hyperparameters in the image processing chain. Therefore,
software application [11] was used to analyze each of the particles, uncertainty associated with image processing is considered negligi-
extracting various metrics for each individual particle such as area, ble compared to the uncertainties shown in Fig. 3. Finally, the
Feret diameter, and roundness. To avoid uncertainty induced by calibration uncertainty (ucal) of SEM measurements in terms of
stitching, non-overlapping images were taken and any particle frequency (% by volume) with respect to each bin was calculated by
touching the edge of a frame was not considered. Since larger summation of all uncertainty components (Fig. 3) in quadrature.
particles have a higher probability of being vetoed by the edge-
exclusion algorithm, the measured PSD may be significantly
3. Results and discussion
skewed. This bias was accounted for via a correction as shown in
Eq. (5), where Pi is the probability of a particle i having a horizontal
SEM results were considered as the calibrated artifact (i.e.,
Feret diameter of xF1 and a vertical Feret diameter of xF2 in a
reference) to determine error (bias) and combined standard
rectangular measurement frame of size Z1 by Z2 [9]. In this
uncertainty associated with DIA measurement results for each bin.
application, there was no preferential orientation found, and
Eq. (5) was therefore simplified to use the same metric for xF1 and
3.1. Results from a DIA instrument
xF2. The population of each size category is divided by its calculated
probability, Pi, where
Powder samples riffled from the same batch were measured in
ðZ 1  xF1 ÞðZ 2  xF2 Þ the DIA instrument taking over 275 images per second with 20 mm
Pi ¼ : ð5Þ field of view and 1 mm resolution, while the particles were
Z1 Z2
dispersed within the measurement volume. Two types of powder
dispersion methods were used in the DIA instrument: dry
2.4. SEM calibration uncertainty dispersion, using a vibratory feeding system passing the particles
through the viewport by a combination of gravity and pressurized
To eliminate systematic error in the SEM image processing, we
used a reference material sample for pixel size (magnification)
calibration [12]. Further, we considered a few relevant sources of
uncertainty in our SEM measurements such as those related to (a)
choice of pixel size, (b) focus and, (c) unevenness of the sample. To
quantify these uncertainties, the same region of dispersed powder
(5350 particles) was analyzed using two different values of these
parameters, chosen over a range that exaggerates the practical
limits for SEM imaging of these particles. Specifically, we acquired
SEM images with pixel sizes of 0.518 mm and 0.259 mm and focus
lengths of 9.8 mm and 10.6 mm. To estimate the effect of
unevenness of the sample, we acquired SEM images at stage
height Z = 9.8 mm and Z = 8.8 mm. We compared the two frequency
distributions resulting from each set of parameters. The differences

Fig. 4. PSD results from a DIA instrument using samples from the same batch. (a)
average results from the dry and wet module and, (b) standard uncertainty in
frequency (% by volume) results from both modules.

air, and wet dispersion, using recirculating liquid suspension of

particles in the viewport. Fig. 4(a) shows the PSD, which is the
average of three repetitions each of dry and wet measurements.
Fig. 4(b) shows standard uncertainties (umeas), in frequency (% by
Fig. 3. Standard uncertainty from each of the considered three sources in SEM volume), associated with the measurement procedure (repeat-
measurements. ability) of DIA with respect to each bin [7].
534 J.G. Whiting et al. / CIRP Annals - Manufacturing Technology 68 (2019) 531–534

DIA results were further analyzed by calculating the difference Finally, the percentiles were calculated using the unique
between each dispersion type (wet and dry) for each ECAD bin of combination of randomly selected frequency values. Since it was
1 mm width. The results are shown in Fig. 4(a). While both assumed that the uncertainty is uniformly distributed, the range
dispersion methods produced reasonable repeatability (an average between the nominal percentile value and the highest percentile,
standard uncertainty of frequencies across all size bins was 0.02% which includes the randomly generated frequencies, is recorded
for wet dispersion and 0.06% for dry dispersion), the wet module over all iterations. This is the uncertainty of the respective
clearly provided a more repeatable measurement, as evident from percentiles as shown in Table 1.
Fig. 4(b), especially for the 25 mm to 35 mm and 42 mm to 50 mm
Table 1
size ranges, where the dry module’s repeatability was worst. ¼ 2) uncertainty and systematic error in
Summary of expanded (k
Investigation of the causes of this increase in variability at these volume fraction percentiles of ECAD for both DIA dispersion
sizes (e.g., drag induced velocity variability, agglomeration of methods.
certain sizes) is a subject of a future study.
D10 (mm) D50 (mm) D90 (mm)

3.2. Uncertainty analysis of DIA measurements SEM 21.77 33.51 50.86

Sys. error  2uDIA [Wet] 0.31  0.58 1.93  0.42 5.45  0.76
Sys. error  2uDIA [Dry] 0.06  0.58 2.13  0.50 5.04  0.82
The SEM data with calibration uncertainty ucal, wet-DIA data,
and the calculated error with respect to SEM are shown, for
example, in Fig. 5(a). Fig. 5(b) shows the SEM and both DIA data 4. Summary

In this study, we showed a novel methodology of using PSD data

from SEM as a calibrated artifact to investigate uncertainty in DIA-
based measurements. This methodology involves converting uncer-
tainty of size measurements into uncertainty in frequencies and then
back to uncertainty of size corresponding to specific percentiles. A
novel method of sample preparation was developed for SEM imaging
of metal powder. To eliminate the uncertainty associated with the
differences of calibrated artifact and the measured sample, samples
with identical characteristics representing the bulk powder proper-
ties were obtained via well-established riffling methods. Both error
and uncertainty of DIA-based ECAD measurements were calculated
using a Monte Carlo simulation and calibrated SEM data. Despite the
proprietary nature of commercially available DIA systems, we have
demonstrated a useful methodology of PSD measurement uncer-
tainty evaluation based on AM metal powder samples calibrated via a
SEM. Future study will involve reducing uncertainty in SEM
measurements (ucal).
Fig. 5. Error analysis of DIA-dry and DIA-wet module. (a) error in frequency of
particles with respect to particle sizes using wet module and, (b) cumulative
frequency (% by volume) of results from all three methods. Acknowledgements

sets in a cumulative plot with common percentiles noted. As This work was partly performed under the financial assistance
illustrated earlier (see Fig. 4(a)), both dry and wet dispersion award number 70NANB18H280 from NIST. Authors thank Dr. Craig
modules have produced very similar measurement results, and M. Shakarji (NIST) for many useful discussions.
therefore their errors trend similarly. Fig. 5(b) shows the smallest
deviation in the cumulative plot from the calibrated SEM data was
at or below the D10 size, with this error increasing through the D50 References
and reaching a maximum around 45 mm (below the D90 size).
The combined standard uncertainty, uDIA, of DIA-based [1] Slotwinski J, Garboczi E (2015) Metrology Needs for Metal Additive
measurements was calculated based on guidelines provided in Manufacturing Powders. The Journal of the Minerals Metals & Materials Society
(TMS) 67(3):538–543.
Ref. [7] and using Eq. (6).
[2] Jacob G, Brown C, Donmez A (2018) The Influence of Spreading Metal Powders
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi with Different Particle Size Distributions on the Powder Bed Density in Laser-
uDIA ¼ ðucal Þ2 þ ðumeas Þ2 þ ðubias Þ2 þ ðusample Þ2 ; ð6Þ Based Powder Bed Fusion Processes. NIST Advanced Manufacturing Series (AMS)
100-17.
[3] Castellanos A (2005) The Relationship Between Attractive Interparticle Forces
where ucal is the uncertainty associated with the SEM calibrated sample, and Bulk Behavior in Dry and Uncharged Fine Powders. Advances in Physics 54
umeas is the uncertainty associated with the DIA-based measurement (4):263–376.
procedure (see Fig. 4b), ubias is the uncertainty associated with the [4] Garboczi E, et al (2017) Particle Shape Effects on Particle Size Measurement for
Crushed Waste Glass. Advanced Powder Technology 28(2):648–657.
systematic error between SEM and DIA measurement results, and [5] Cipolla R, Blake A (1990) The Dynamic Analysis of Apparent Contours. Proceed-
usample is the uncertainty associated with the differences between ings of the 3rd International Conference on Computer Vision, 616–623.
calibrated sample and the actual measured sample. Since SEM [6] ISO 13322-2 (2006) Particle Size Analysis – Image Analysis Methods – Part 2:
Dynamic Image Analysis Methods.
measurements are used as the calibration of the artifact, ubias is [7] ISO 15530-3 (2014) Geometrical Product Specifications – Coordinate Measuring
calculated as the standard deviation of the mean of DIA measurements. Machines: Technique for Determining the Uncertainty of Measurement - Part 3:
Furthermore, we considered usample as negligible sincewe used the same Use of Calibrated Workpieces or Measurement Standards.
[8] Terence A (1997) Particle Size Measurement - Volume 1: Powder Sampling and
batch of powder for all measurements and sampling was done via a
Particle Size Measurement, Springer, Netherlands. ISBN 978-0-412-72950-8.
rifflingoperation. To calculate the combined standard uncertaintyof DIA [9] ISO 13322-1 (2014) Particle Size Analysis – Image Analysis Methods – Part 1:
(uDIA) of each of the conventionally reported percentiles (i.e., D10, D50, Static Image Analysis Methods.
[10] Kuwahara M, et al (1976) Processing of RI Angiocardiographic Images. Digital
and D90), a Monte Carlo simulation with 500 000 iterations was used.
Processing of Biomedical Images, Springer, Boston, MA187–202.
First, we calculated the uDIA in terms of frequency for each bin. Then, a [11] Schindelin J, Arganda-Carreras I, Frise E, et al (2012) Fiji: an Open-source
random frequency value was selected between the upper and lower Platform for Biological-image Analysis. Nature Methods 9(7):676–682.
limits of each uncertainty band. This was done foreach bin that was used [12] Postek M, et al (2010) Reference Material (RM) 8820: A Versatile New NIST
Standard for Nanometrology. Metrology, Inspection, and Process Control for
in calculating the respective percentile. After each iteration, the bin’s Microlithography XXIV. in Raymond C, (Ed.) Proceedings of SPIE, (7638):
frequency values were normalized to ensure they sum to unity. 76381B-1.