Page 19 - ESAM-1-4

Engineering Science in
Additive Manufacturing Experimental statistics in AM

applications. We therefore recommend the development of However, this only represents about 35% of all sampled
a list of standard experimental statistics practices to be used papers, while most others use poor experimental designs
in AM. What is outlined in this paper is a proposed first such as one factor at a time and single-factor experiments.
step in this direction. For instance, the researcher reporting The first design necessarily excludes interactions from
the results of experiments would mention the type of statistical models, and the latter design only explores a
design used, the purpose, and a justification for the sample single source of variation. Other papers also may not
size used. The steps for selecting the statistical model, as run a formal design at all and only report the results of a
well steps to evaluate the adequacy of that model, should single run or repeated runs of the same design point. Very
be given. Finally, the results of the final regression and few experiments used standard RSM designs, such as a
ANOVA should be shown (t-value, the number of degrees CCD. One design which has been confused for a standard
of freedom, R , p-value, regression coefficients, etc.). RSM design is the group of Taguchi designs, which were
2
Standardization, though, does not necessarily entail the designed for the field of robust parameter design. 55,56
use of more rigorous analyses and designs. Studies using 3D These designs are meant to minimize the effect of noise
printing to replace orthopedic tissue use standard designs factors on a response variable by adjusting the values of
control factors. Despite requiring an excessive number
(mostly single-factor) and analyses (ANOVA and t-test), of runs compared to a similarly sized fractional factorial
which are sufficient for simple hypothesis testing; they are design, the approach does not accommodate potentially
not efficient insofar as they only test one hypothesis at a important control factor interactions. When all factors
57
time, and they may ignore nonlinear effects such as those
modeled by interactions and quadratic terms. Given that are controlled, but some are inadvertently treated as noise
any experimental result is likely the product of multiple factors, important low-level interactions can be missed or
interacting variables, multiple hypotheses should be tested misinterpreted. In addition, if the split-plot nature of the
design is not accounted for in the fitted models, the power
simultaneously using RSM designs to ensure the economic of the test can be lowered. If Taguchi designs are to be used
viability and the robustness of the final result. In addition, outside of their intended context, they should be combined
48
the use of proper experimental techniques did not change with traditional response surface methods to control for
much between 2016 and 2024. Standardization, then, these shortcomings. However, in our literature review,
58
should also include higher standards for experimental we found that Taguchi designs were used in place of other
statistics.
designs, which is problematic. Instead, practitioners could
Still, the use of even these basic designs and analyses be using CCDs, Box–Behnken designs, and computer-
likely helps buffer the medical field against the crisis of generated designs, which optimize various selection
replication, as improper designs can artificially inflate criteria such as D-efficiency, when evaluating a response
false positive rates. In medicine, unreproducible results surface. Full factorial and fractional factorial designs
49
are unacceptable given the potential cost of human life, should be used for screening factors.
but they should also not be tolerated in any scientific field. Given that trends in the use of these analyses and
Despite this, multiple engineering fields have been slow designs appear to remain unchanged over time, there is a
to adopt DOE methods, 50,51 potentially contributing to a need to re-evaluate their application in AM broadly and in
replicability crisis in the engineering field akin to that seen PBF-LB/M specifically. In section 2 and in the companion
in psychology. Steps should also be taken to ensure the manuscript to this guide, we present a standardized
52
26
validity of the statistical model, as model misspecification method of applying DOE methods to AM and PBF-LB/M
can result in incorrect inferences, further fueling issues with problems, and discuss the minimum requirements for
reproducibility. 53,54 In addition, model misspecification reporting these methods in a manuscript or report. If
caused by multicollinearity can also reduce the precision these recommendations are implemented, they have the
54
of these models and inflate type I error rates. At the very potential to accelerate progress in the field by reducing
least, we encourage AM practitioners to adopt the practices sample sizes, removing communication barriers imposed
48
used in orthopedic engineering to avoid these problematic by imprecise experimental descriptions, and enhancing
outcomes. the reproducibility of research. These recommendations
59
That said, we still advocate for stronger experimental could also be used in the context of a round-robin study to
designs, which can reduce the cost of experimentation in distribute the costs of the experiments across laboratories.
60
numerous ways. 23,25,48 The most common design choice in This guide is useful for both practitioners and editors. Peer-
AM studies was the generalized full factorial design. While reviewed journals, especially those specialized in AM, can
suitable in many cases, it is often information-inefficient request proper use and reporting of experimental statistics,
compared to fractional factorial and other related designs. but they can also advocate for better designs. Adoption of

Volume 1 Issue 4 (2025) 13 doi: 10.36922/ESAM025340021