PSI - Issue 75

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

95

2

assessment and are still the base of common recommendations (Berger et al. , 2006; Eurocode 3: Design of steel structures -Part 1-9: Fatigue, 1993-1-9:2005 , 2009; Fiedler et al. , 2019; Baumgartner, Hobbacher and Rennert, 2020; Rennert et al. , 2020). Pertaining to the subject matter of this study, (Agrawal et al. , 2014) demonstrated the feasibility of predicting the fatigue strength of steels from chemical composition and processing parameters, specifically using the data from (Furuya et al. , 2019) restricted to rotating bending experiments at room temperature. In this regard, they also launched an online tool (Agrawal and Choudhary, 2018). (Wang et al. , 2023) expanded this approach with physics-informed methods, while (Liu et al. , 2023) concentrated on result interpretability. Other collections of fatigue databases do exist, such as those in the context of welds of construction steels (Bartsch et al. , 2020) or offshore structures (Zhao, 2021) which could potentially be utilized as training data for machine learning (ML) predictions of fatigue strength for other specific use scenarios. ML can also aid in the improved design of experiments, for instance, to recommend suitable stress levels with high statistical significance during the experimental campaign to derive an S-N/Wöhler curve (Weichert et al. , 2022). The successful implementation of ML relies on an appropriate dataset for training, which is typically still supplied as a data frame or spreadsheet, featuring several categories at a flat hierarchy level. In previous work, the conceptional application of a ML model (Fliegener et al. , 2023; Fliegener, J. Rosenberger, M. Luke, J. Domínguez, J. Morgado, H.U. Kobialka, T. Kraft, 2024) is used to predict the fatigue strength of a wide range of steels based on an extensive database. In this database, a high number of around 22.000 individual fatigue tests on simple steel specimens is available. The predictions, however, was limited to typical SN-curve parameter like slope or alternating strength (fatigue strength at =−1 ). The reached 2 scores range between 0.7 and 0.8. The database was also used to predict the fatigue strength of welded joints based on the a ML approach trained on conventional steel specimen (transfer learning) (Schubnell et al. , 2025). In this work, the mentioned approach is extended to a direct prediction of the fatigue life of single specimen by ML. The training data space, ML-approach and target parameter (fatigue life or fatigue strength) was varied. In section 2 the different approaches are introduced, while in section 3 the data space and in section 4 the different ML models are described. The results are displayed in section 5 and the conclusions are given in section 6. The aim of this work is to identify the limitations of data driven approaches for fatigue assessment based on an extensive database of steels.

Nomenclature HV

Failure probability Load cycles 10 6 load cycles stress at 10 6 LC and ൌͷͲΨ ML Machine Learning Artificial neural network ANN

Average roughness Vickers hardness Stress amplitude Stress ratio SCF

[µm] [HV]

[%]

[-]

Stress concentration factor

[-]

[µm] [MPa]

2

[MPa]

[-]

RSME

Root square mean error

[MPa]

R2-score

[-]

TL

Transfer Learning

2. Principle strategies for fatigue prediction based on ML Two principal approaches or strategies for fatigue assessment were used in this work illustrated in Figure 1. As mentioned in this work the target parameter is the fatigue life in number of cycles. The fatigue life could either be determined if the SN-curve parameter is known (fatigue strength and slope) by equation (1), illustrated in Figure 1 (a) or the fatigue life could be determined directly if the stress amplitude is used as an additional input parameter, see Figure 1 (b). The input parameter are the surface hardness in [HV], stress concentration factor (SCF), surface roughness (in this work the values according to DIN 4760:1997 in [µm] was used), the stress ratio the load type (tension, bending, torsion, rotary bending), the testing temperature in [k] and the chemical composition in [%]. The last parameter again contain a range of single element portions of {Cr, C, Ni, Cu, Si, Mo, Mn, N, S, Al, P, V, Co}. = ×( ) 1 (1)