A 2.830 (MIT) study quantifying how key BAAM process parameters influence single-bead geometry using a structured DOE,
ImageJ measurement, and statistical analysis (ANOVA, nested ANOVA, regression).
We analyzed how nozzle Z-offset, extrusion RPM, and feed rate affect dimensional accuracy
in BAAM single-bead printing using a fractional DOE, then quantified main effects/interactions using ANOVA and regression.
DATASET & CONTROLLED CONDITIONS
Dataset included 45 single-bead images (we focused on single-bead only).
Constants held fixed: melt temperature 277°C, return time 272s, 10 layers/part, nozzle diameter 0.4".
Measurements were extracted from cross-section images using ImageJ.
MEASUREMENT PROCEDURE
We targeted five geometric outputs: layer height, layer width, overall part height, overall part width, and contact width.
For consistency, layer-level metrics were averaged across all 10 layers; overall height/width were averaged across three locations.
DESIGN OF EXPERIMENTS
A full 3³ factorial would require 27 treatments. To reduce burden, we used a 9-treatment, 3-level fractional design
equivalent to an L9 Taguchi orthogonal array (strength 2), with 5 replicates per treatment, focusing analysis on
main effects and two-way interactions.
Run + Moving Range charts used to visualize repeatability and detect drift / special-cause behavior.
Experimental structure intentionally induced deterministic changes via parameter blocks.
Effect estimation
One-way ANOVA per response to test treatment differences.
Nested ANOVA (variance components) to compare within-treatment noise vs between-treatment effects.
OLS regression using main effects + two-way interactions.
KEY RESULTS
Across outputs, replicate-to-replicate variability often dominated treatment differences; nested ANOVA attributed most variance to within-treatment noise.
Regression models were not significant overall, despite a few individually significant coefficients—consistent with the nested ANOVA conclusion.
Directional/physical interpretation still emerged: nozzle Z-offset (LH) behaved as the dominant lever, feed rate was second-order, and RPM had comparatively smaller effects.
LIMITATIONS & WHAT WE’D DO NEXT
Within the safe operating window tested, the physical response of the printer and measurement variability reduced contrast between treatments.
We recommended widening parameter excitation, improving replication strategy (randomized true replicates), and adding factor levels / broader coverage
to increase statistical power and capture non-linear behavior.
