PSI - Issue 75

ScienceDirect Available online at www.sciencedirect.com ScienceDirect Structural Integrity Procedia (2025) 000 – 000 Available online at www.sciencedirect.com ScienceDirect Structural Integrity Procedia (2025) 000 – 000 Available online at www.sciencedirect.com Procedia Structural Integrity 75 (2025) 94–101

www.elsevier.com/locate/procedia

www.elsevier.com/locate/procedia

2452-3216 © 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under the responsibility of Dr Fabien Lefebvre with at least 2 reviewers per paper 10.1016/j.prostr.2025.11.011 2452-3216 © 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of the Fatigue Design 2025 organizers 1. Instroduction Data-driven or machine learning (ML) approaches are effective if large data sets with a high variance are available, and no accurate or comprehensive physical description of the phenomena exist (Ross, 1993; Lee, Almond and Harris, 1999; Uygur et al. , 2014). For the fatigue assessment of materials or industrial and components parts this is the case. Numerous models for the description of the fatigue phenomena exist but usually no model covers the complete period of fatigue, starting in an atomistic scale to cyclic slip, crack nucleation, crack propagation in different scales (microstructural- and mechanical short cracks and mechanical long cracks) to the final fracture (Haibach, 2006; Radaj and Vormwald, 2007). Cost-extensive experiments and fatigue tests are still the standard procedure of fatigue 2452-3216 © 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of the Fatigue Design 2025 organizers 1. Instroduction Data-driven or machine learning (ML) approaches are effective if large data sets with a high variance are available, and no accurate or comprehensive physical description of the phenomena exist (Ross, 1993; Lee, Almond and Harris, 1999; Uygur et al. , 2014). For the fatigue assessment of materials or industrial and components parts this is the case. Numerous models for the description of the fatigue phenomena exist but usually no model covers the complete period of fatigue, starting in an atomistic scale to cyclic slip, crack nucleation, crack propagation in different scales (microstructural- and mechanical short cracks and mechanical long cracks) to the final fracture (Haibach, 2006; Radaj and Vormwald, 2007). Cost-extensive experiments and fatigue tests are still the standard procedure of fatigue © 2025 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under the responsibility of Dr Fabien Lefebvre with at least 2 reviewers per paper Abstract Machine learning (ML) approaches gain more and more importance for fatigue assessment of materials and industrial parts. In this work an extensive database of more than 22.000 single fatigue test and 1100 fatigue test series (SN curves) of different steels are used to build a generalized approach for the fatigue prediction based on machine learning. For this, different strategies are used: First, on SN-curve level, the fatigue life assessment based on SN curves, where the SN-curve parameters (slope, fatigue strength) were determined by ML and used for the fatigue life prediction later; and second, the fatigue life prediction based on specimens, where the characteristics of single specimen of the fatigue tests series (stress amplitude, roughness, hardness, … ) are used (specimen level). Different ML approaches like an artificial neural network (ANN) or random forest approach are used. A higher accuracy of the direct fatigue life prediction is shown. Slightly higher accuracy was determined using an ANN. This work shows the limitation using mainly commercial, older data sources for ML-based fatigue assessment with a certain degree of inconsistency that affect the prediction accuracy of this approaches. Fatigue Design 2025 (FatDes 2025) Principle strategies for the fatigue assessment of steels based on machine learning approaches Jan Schubnell a *, Anastasiia Danchenko a , Johannes Rosenberger a , Sascha Fliegener a , Michael Luke a a Fraunhofer Institute for Mechanics of Materials, Woehlerstr. 11, Freiburg, Germany Abstract Machine learning (ML) approaches gain more and more importance for fatigue assessment of materials and industrial parts. In this work an extensive database of more than 22.000 single fatigue test and 1100 fatigue test series (SN curves) of different steels are used to build a generalized approach for the fatigue prediction based on machine learning. For this, different strategies are used: First, on SN-curve level, the fatigue life assessment based on SN curves, where the SN-curve parameters (slope, fatigue strength) were determined by ML and used for the fatigue life prediction later; and second, the fatigue life prediction based on specimens, where the characteristics of single specimen of the fatigue tests series (stress amplitude, roughness, hardness, … ) are used (specimen level). Different ML approaches like an artificial neural network (ANN) or random forest approach are used. A higher accuracy of the direct fatigue life prediction is shown. Slightly higher accuracy was determined using an ANN. This work shows the limitation using mainly commercial, older data sources for ML-based fatigue assessment with a certain degree of inconsistency that affect the prediction accuracy of this approaches. Keywords: Fatigue, steels, machine learning, random forrest, articifical neural network, statistical evaluation Fatigue Design 2025 (FatDes 2025) Principle strategies for the fatigue assessment of steels based on machine learning approaches Jan Schubnell a *, Anastasiia Danchenko a , Johannes Rosenberger a , Sascha Fliegener a , Michael Luke a a Fraunhofer Institute for Mechanics of Materials, Woehlerstr. 11, Freiburg, Germany Keywords: Fatigue, steels, machine learning, random forrest, articifical neural network, statistical evaluation