PSI - Issue 57

Jan Schubnell et al. / Procedia Structural Integrity 57 (2024) 112–120 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

117

6

The results are illustrated as scatter plots in Figure 5. For better interpretability a diagonal is shown which corresponds to the perfect prediction in which the predicted SCF corresponds exactly to the target SCF. As can be seen both PointNet++ and 2DLaserNet perform equally well with predictions lying close to the ideal diagonal. Only one outlier is visible. The R2 score in this case corresponds to 0.947 for PointNet++ and 0.97 for 2DLasernet (maximum values of 1.0).

(a)

(b)

PointNet++

2DLaserNet

Fig. 4. Scatter plots for comparison of the stress concentration factor taken from FEM and the stress concentration factor determined by ANN model for the comparison of (a) PointNet++ and (b) 2DLaserNet trained by virtual 2D-profiles 3.6. Comparison of ANN and other SCF-solutions To evaluate the efficiency of the proposed method for the determination of SCF a comparison was made with analytical SCF solutions based on geometrical parameters SCF( , , , ), given by Anthes et. al. (Anthes, Köttgen and Seeger, 1993), Rainer (Rainer, 1978) and Kiyak (Kiyak, Madia and Zerbst, 2016), summarized in Table 1. Table 1. Analytical SCF solutions

Range ( ) mm Range ( ) [°] Range ( )

Range ( )

Reference

Formulae

(1 − 0.156 ∗ ( ) 0.207 ) ∗( 1 +(0.181 +1.207 ∗ sin( )−1.737 ∗ sin 2 ( ) + 0.689 ∗ sin 3 ( ))∗ ( ) 0.2919+0.3491∗sin( +3.283) ) 0≤ ≤200 0 - 90

Anthes et. al. (Anthes, Köttgen and Seeger, 1993) Rainer et. al. (Rainer, 1978)

0≤ ≤200 - 0≤ ≤400 -

0≤ ≤400 45

2.25 +0.2∗ 2

−0.5

1 + (( 0.4 ) 0.66 + 3.8 ∗ ( 1 + 2 2 √ 2 )

( 2 + ) ( ) 1.33 )

1 + 1.1399 ∗ (ℎ ) 0.2062 ∗ ∗ 1.067 ∗ −1.6775 ∗ ∗ ( ) −0.295∗ ∗ (0.021 + ) −0.4711

0.1 ≤ ≤ 4 10 ≤ ≤ 60 10 mm

1

Kiyak et. al. (Kiyak, Madia and Zerbst, 2016)

4.6 mm

The input parameters for these analytical SCF solutions are the weld toe radius , flank angle , plate thickness and in some cases the weld width . These geometrical parameters are already given in the case of virtual 2D-profiles, see Figure 2 (a). The results are displayed in Figure 5. As shown, the relatively new SCF solution by Kiyak et. al. (Kiyak, Madia and Zerbst, 2016) and the solution of Anthes et. al. (Anthes, Köttgen and Seeger, 1993) show the lowest