PSI - Issue 75

Jan Schubnell et al. / Procedia Structural Integrity 75 (2025) 94–101 Schubnell/ Structural Integrity Procedia (2025)

97

4

Table 1. An example of a table. Characteristics of the database on fatigue behavior of engineering steels (Fliegener, J. Rosenberger, M. Luke, J. Domínguez, J. Morgado, H.U. Kobialka, T. Kraft, 2024) Data sources DaBef , Boller&Seeger, NIMS datasheets, JSMS databook, misc. literature see Fliegener et. al. (2024) Scope 22422 data points on specimen level, 1144 on series level, 111 materials Degree of completeness Degree of completeness of the main categories (series with value specified divided by total number of series): Fatigue strength (841 series, 74 %), Hardness HV (1002 series, 88 %), Tensile strength Rm (1007 series, 88 %), Roughness Rz (740 series, 65 %), Stress concentration factor Kt (1124 series, 98 %), Loading type (1110 series, 97 %), R ratio (1091 series, 95 %), cross section geometry (845, 74 %)

Steel groups

According to DIN 10027-2, JIS, AISI, SAE: - Alloyed construction, engineering and pressure vessel steels - Construction steels - Heat-resisting and stainless steels - Low-alloyed construction, engineering and pressure vessel steels - Other steels - Roller bearing steels - Spring steels - Tool steels - Unalloyed construction, engineering and pressure vessel steels

Table 2. Data spaces used for ML training Data space No. of SN curves No. of specimen

Materials

Description

DS1 DS2 DS3 DS4

1152 1152

111 materials 111 materials 49 materials 49 materials

22527 12030

Include all data

No runouts, fatigue life < 10 7 No runouts, fatigue life < 2 × 10 6 , < 1000 MPa No runouts, fatigue life < 2 × 10 6 , < 1000 MPa, specimen averaged according to equation (2) for specimen at equal No runouts, 10 4 < fatigue life < 2 × 10 6 , < 1000 MPa, specimen averaged according to equation (2) for specimen at equal

477 305

8892 3092

DS5 DS6

305 109

2709 1095

49 materials

100Cr6

4. Application of ML approach Two different machine learning approaches was used in this work, illustrated in Figure 2. First, the decision tree based Random Forrest approach was used according to former work (Fliegener, J. Rosenberger, M. Luke, J. Domínguez, J. Morgado, H.U. Kobialka, T. Kraft, 2024), illustrated for an hardness prediction in Figure 2 (a). Second, a shall artificial neural network is used, shown in Figure 2 (b). A random forest regressor is a machine learning algorithm that falls under the ensemble learning category, specifically within the bagging methods family (Breiman, 2001). For each model estimator, a decision tree is grown. Each tree is trained on a distinct bootstrapped subset of the training data to enhance diversity among the decision trees. This amplifies the overall generalizability of the random forest model and reduces the risk of overfitting. In the bootstrapping method (Feron, 1979), the fundamental assumption is that the random sample is "representative" of the population it was drawn from. This population is now substituted by the sample. New samples are generated by repeated sampling with replacement. When splitting each node during the tree's construction, the best split is identified through an exhaustive search of the values of either all input features or a random subset of size max features. According to (Pedregosa, 2011), considering all features is an empirically sound default for regression tasks. During prediction, each tree in the forest yields an output. For regression tasks, the final prediction is typically the average of the outputs from all the trees.