PSI - Issue 79

Available online at www.sciencedirect.com

ScienceDirect

Procedia Structural Integrity 79 (2026) 190–197

© 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 IGF28 - MedFract3 organizers Keywords: Constitutive model; Machine learning; Data augmentation; Scarce data This study explores two di ff erent data augmentation approaches to generate artificial data for the training process of machine learned constitutive models. The cyclic Ramberg-Osgood model was employed by predicting the stress amplitude at 1% strain amplitude in the materials. A baseline model using 96 materials in the training process was first defined. Ten materials out of the 96 were sampled to be augmented. The first augmentation approach was to perform a linear numerical scaling of sampled materials to generate artificial copies. However, this approach lack the real physical behaviour of the material. Thus, to account for the physical behaviour, an augmentation based on empirical statistical distributions of the material behaviour of the sampled materials was made. This generates improved artificial data with a reliable outcome. Hence, a robust, accurate and e ffi cient training procedure was obtained compared to used experimental data. 28th International Conference on Fracture and Structural Integrity - 3rd Mediterranean Conference on Fracture and Structural Integrity Data augmentation of scarce experimental data for constitutive models Daniel Leidermark a, ∗ , Robert Eriksson a a Division of Solid Mechanics, Linko¨ping University, 58183 Linko¨ping, Sweden Abstract

1. Introduction

Accurate predictions of the stress-strain state in a component during the design phase are of utmost importance for structural integrity and load-carrying capability. With the evolution of artificial intelligence, or machine learning (ML), possibilities now arise to completely replace the need for traditional constitutive models, i.e. models based on empirical relations and engineering assumptions. In general, the accuracy and the ability to make predictions beyond given training limits for an ML model is defined based on the number of available training data. This usually implies thousands to hundred of thousands data points. However, this is not the case when it comes to available experimental tests for training of constitutive models, where commonly very scarce number of unique mechanical tests are available, usually 10 to 20. This relates to that experimental testing is time consuming and expensive, and as a consequence will, with high probability, give a poorly trained ML model. A remedy for this can be data augmentation, which is a technique to artificially expand the present number of data points, to increase

∗ Corresponding author. Tel.: + 46-013-282791. E-mail address: daniel.leidermark@liu.se

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 IGF28 - MedFract3 organizers 10.1016/j.prostr.2025.12.324