Page 60 - MSAM-4-3

Materials Science in Additive Manufacturing Bead geometry prediction in laser-arc AM

bead dimensions is both intricate and non-linear, calling chemical compositions of the substrate and the wire are
for further intensive study. Because individual regression provided in Table 3.
models are constrained by their respective hypothesis
spaces, no single model can guarantee an optimal result; 2.2. Experiment design
therefore, ensemble schemes that combine multiple Within the WAAM process, the weld bead dimensions are
models are generally employed to enhance prediction strongly affected by the choice of process parameters. An
33
accuracy. Put differently, simple ML or neural-network increase in wire feed speed leads to greater weld bead width
approaches alone may not suffice, whereas more elaborate and height. A larger arc length adjustment broadens the
34
ensemble architectures are able to offer superior predictive bead width, whereas a higher pulse correction decreases
35
accuracy. In addition, incorporating explainable-analysis its height. Within LAHAM, an optimal laser power level
methods enables researchers to grasp data features more supports geometric uniformity, whereas overly high power
thoroughly and to probe the effect of process parameters causes size variability. Accordingly, this study concentrates
36
on bead-geometry predictions. These analytic approaches, on the influence of wire feed speed (v ), welding speed (v ),
t
w
once validated, support investigators in deepening their arc length correction (l), pulse correction (f), and laser
data understanding and hence in assessing more precisely power (p) on weld bead width (W) and height (H).
the impact of process parameters on predictive outcomes. In Figure 3, the full factorial design with three factors
Herein, we present a particle swarm optimization at three levels and the Box–Behnken design are depicted
37
(PSO)-based ensemble predication (PSO-EP) that schematically. The Box–Behnken design, as opposed
combines four base models – Gaussian process regression to the full factorial design, prevents extreme condition
(GPR), SVR, artificial neural networks (ANN), and combinations and efficiently captures second-order effects
extreme learning machines (ELM) – with PSO employed to with a reduced number of experiments, and was therefore
calibrate their respective weights. The predictive capability employed in this work. First, this study followed the process
38
of PSO-EP is benchmarked against single base learners and window recommended in reference and employed single-
alternative ensemble methods, and the results verify the factor screening experiments to define the valid ranges of
superiority of the proposed approach. In addition, based each influencing factor, encoding their upper and lower
on Shapley theory, an interpretability study is carried out bounds as +1 and –1, respectively; the selected key factors
in which visual tools, including feature-importance and and their symbols are detailed in Table 4. Table 5 displays
sample-distribution charts, afford deeper insights into how the 46 coded experimental conditions and outcomes.
individual features and samples sway the predictive results. Table 6 comprises 20 randomly sampled process parameter
configurations and associated weld geometry metrics,
2. Materials and methods used as the validation and test datasets. To assess the
reproducibility of the experimental dataset, six repeated
2.1. Experimental setup trials were conducted under the central-point parameter
39
The system utilized in this study is depicted in Figure 1. configuration, and the coefficient of variation (CV) was
In the system, the composite heat source is composed employed for evaluation. The analysis revealed CV values
of a welding machine (Fronius CMT Advanced 4000R, of 3.25% for bead width and 2.74% for bead height –
Austria) and a fiber laser system (Raycus RFL-C3300W, both markedly lower than the commonly accepted 10%
40
China). The fiber laser system consists of a laser source, an benchmark – providing strong evidence of high data
output head, and a cooling unit, with the relative position consistency. As illustrated in Figure 4, all the weld beads
of the laser output head and the welding gun is illustrated were well-formed and defect-free. The overall experimental
in Figure 1. Line structured light (Gocator 2430, Canada) count was determined based on Equation I.
was utilized to obtain the point cloud data of the weld bead N = 2q(q−1)+C (I)
morphology. The procedure of laser-scanning the actual 0
weld bead morphology and generating the point cloud is where q is the number of experimental parameters, and
depicted in Figure 2. The motion platform used is a CNC C denotes the number of repetitions needed to minimize
0
machine tool (Fana FA2818HG, China). The deposition errors arising from environmental and human factors. In
material is ER2319 aluminum alloy (1.2 mm diameter), this paper, q is 5, and C is 6.
0
and the base material is 2219 aluminum alloy. During the
manufacturing process, the welding mode was cold metal 2.3. Particle swarm optimization-based ensemble
transfer pulse advance (CMT-PADV), with the laser CMT prediction
process parameters listed in Table 2. Before the experiment, In the proposed ensemble forecasting framework, the
the substrate was processed and cleaned with acetone. The four base learners – ANN, GPR, SVR, and ELM – are first

Volume 4 Issue 3 (2025) 3 doi: 10.36922/MSAM025220036