Page 11 - ESAM-1-4

Engineering Science in
Additive Manufacturing Experimental statistics in AM

candidate list of papers was created for PBF using the from the text alone, the paper was abandoned and replaced
search terms “Powder Bed Fusion” AND “Experiment” OR with a new randomly selected paper. If other aspects of the
“Machine-Learning” AND “Selective Laser Melting” OR design were not stated (such as whether the experiment
“Selective Laser Sintering,” and its relevance was checked was performed in random order), it was assumed that this
as well. We included a search term for “machine learning” aspect of the design was absent from the study. If there
to capture cases where machine learning models were used were multiple experiments performed, then data would
to evaluate experimental results instead of or in addition to only be collected on the first experiment mentioned in the
traditional linear models. text. Figure 2 shows the workflow of the review, and Table 3
Next, we randomly sampled up to a minimum of 12 gives a brief overview of each of the topics we covered in
papers from each year from 2016 to 2024 in Scopus for both the review. Each of these topics is discussed in more depth
26
AM and PBF-LB/M, increasing the number of samples in the companion manuscript.
every year to reflect the acceleration in publication rates 3.2. Summary results of literature review
in the field (Table 2). We restricted our analysis to papers
published between 2016 and 2024 because this period Overall, the data reveal no significant temporal trends in
captures the rapid expansion of AM research following the use of design types or analytical methods for either
the release of key standards, such as ISO/ASTM 52901 AM or PBF-LB/M. However, the two approaches exhibit
on requirements for purchased AM parts and ISO/ASTM differences in the specific designs and analyses they
52915 on standardized AMF file formats, 27,28 ensuring employ. First, we find the proportion of papers each year
that temporal trends reflect practices developed under a that utilize different aspects of DOE and ANOVAs for all
modernized and increasingly standardized framework. AM (Figure 3, left panel) and PBF-LB/M (Figure 3, right
Before 2016, papers on this topic were relatively scarce. panel). We then find the proportion of papers each year
Each randomly selected paper was vetted to make sure that perform different data analysis (Figure 4) and use
it was an experimental paper (as opposed to a review or different design types (Figure 5). We then fit a LOESS
29
theoretical paper). If a paper did not meet this requirement, regression to each category to measure trends over time.
it was excluded and replaced with a new random paper In each case, these regressions tend to either remain flat
from that year. over time or fluctuate wildly, so no significant trends can
be predicted from this analysis. Finally, we categorize
Each paper was assessed on the type of data analysis, some designs (full factorial, fractional factorial, Taguchi,
the type of experimental design, whether blocking was central composite, and randomly distributed factor levels)
present, whether the run order was randomized, whether as being more sophisticated than alternative designs and
model adequacy checks were performed, whether plot the proportion of this group of designs over time. We
model selection took place, and whether there was some then performed a linear regression, which includes this
justification for the sample size, and finally, the sample proportion as its response variable, while year and field type
size itself was recorded. If aspects of the design were not are predictor variables (Figure 5; linear model: p-values for
explicitly stated, attempts were made to estimate values all effects >0.05, p-value for the model = 0.5866, multiple
such as sample size from figures or attached datasets. If the R = 0.1249). Neither their main effects nor interactions are
2
sample size, design, or analysis could not be determined
significant, despite that the fit was adequate (residual plots
were checked). This means that designs are not becoming
Table 2. Counts of manuscripts sampled per year across the
different manuscript types increasingly sophisticated over time.
We also investigated differences in experimental
Year All Powder Orthopedic practices between AM and PBF-LB/M in aggregate. First,
AM bed fusion engineering
2016 12 15 10 there are no differences in the proportion of good practices
used by AM or PBF-LB/M papers (Figure 6A; all sample-
2017 13 16 0 size corrected proportion tests have p>0.05), nor are there
2018 14 17 0 differences in sample sizes (Figure 6B; Wilcoxon test
2019 15 18 0 p>0.05). In addition, there is no difference in the proportion
2020 16 19 0 of analyses or experiments present in AM manuscripts
2021 17 20 0 than in PBF-LB/M manuscripts. To test this, we performed
2022 18 21 0 a binomial generalized linear model (GLM) with the main
effects of analysis type and manuscript type, as well as an
2023 19 22 0 interaction between these two factors. This interaction
2024 18 21 10 between PBF-LB/M and the type of analysis has a p>0.05

Volume 1 Issue 4 (2025) 5 doi: 10.36922/ESAM025340021