PSI - Issue 75

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

96

3

(a) Prediction on SN-curve level

Fliegener et. al. (2024)

Input Parameters: Hardness [HV] SCF [-] Roughness [µm] Stress ratio R [-] Load type [-] Temperature T [k] Chem. Composition[El. / %]

ML Random Forrest / ANN

SN-curve parameter Fatigue strength* [MPa] Slope [-] Scatter range [-]

Single specimen fatigue life Number of cylcles [-]

(b) Prection of specimen level

Input Parameters: Hardness [HV] SCF [-] Roughness [µm] Stress ratio R [-] Load type [-] Temperature T [k] Chem. Composition[El. / %] Stress amplitude [MPa]

ML Random Forrest / ANN

Single specimen fatigue life Number of cylcles [-]

Fig. 1. Approaches for fatigue life determination by ML

3. Data spaces The used database comprises 111 material types, 1144 test series, and 22422 fatigue experiments, which have been assembled from literature data and previous projects. A fatigue experiment is defined as a single specimen that either failed at a certain number of cycles or reached the defined fatigue limit. Conversely, a test series is used to determine an S-N curve, which consists of multiple fatigue experiments. The data was systematically collected, digitised, post-processed, and curated. The categories and parameters taken into consideration are detailed described by (Fliegener, J. Rosenberger, M. Luke, J. Domínguez, J. Morgado, H.U. Kobialka, T. Kraft, 2024). Most parameters describing the S-N curve properties, loading type, composition, condition, and geometry are defined at the series level. In contrast, the only data acquired at the specimen level is the stress amplitude, number of cycles, and classification, determining whether a specimen was a runout. For evaluation purposes, a hierarchical structure of steel categories was developed in accordance with various standards such as DIN EN 10027-2, JIS, AISI, and SAE. If a test series corresponding to these standards exists in the database, a steel group was added. An overview of the database, including the data sources and the steel groups, is provided in Table 2. This table also shows the main categories and their degree of completeness, indicating the percentage of series in which a value for a specific parameter (e.g., surface hardness) is provided. In some instances, the literature data was incomplete, resulting in the degree of completeness being less than 100%. In other cases, the respective value could be determined through conversion processes (e.g., deriving hardness values from tensile strength). More details about data curation are also provided. Given that various data sources were considered for data acquisition (as shown in Table 1), the initial data was strongly heterogeneous. To enable a uniform and consistent evaluation of the fatigue strength for the complete database, the raw data (i.e., stress amplitude and number of cycles for each specimen) was collected. n contradiction to former studies (Fliegener, J. Rosenberger, M. Luke, J. Domínguez, J. Morgado, H.U. Kobialka, T. Kraft, 2024; Schubnell et al. , 2025) the database was statistically evaluated according to DIN 50100:2022 ( DIN , 2022) to assure a high degree of consistency. According to this standard the slope and fatigue strength based on the current FKM-guideline (Stress amplitude at = 50% at = 1 × 10 6 load cycles were evaluated). For this work different data spaces (DC1 to DC6) are used for the training of the ML approaches, see Table 2. To increase the consistency of the data the data space was stepwise reduced and simplified from DC1 to DC6. For DC4, DC5 and DC6 the number of cycles given for different specimen was averaged if an equal stress amplitude was used in the fatigue tests: (̅̅̅̅ )= 1 ∑ ( ) = 1 for all = (2)