Universal characteristics of local strain fields for creep failure prediction

Material creep, defined as time-dependent strain accumulation under constant loading, can result in severe deformation and eventual component failure, posing a significant engineering challenge. Therefore, the possibility of early prediction of creep behavior is highly desirable. The objective of this study is to propose a robust method for predicting creep failure. To this end, we investigate the creep behavior of paper samples (quasi-brittle fiber composites) used as a model material, subjected to constant uniaxial tensile loads. Local strain fields are obtained through Digital Image Correlation and analyzed using dimensionality reduction techniques, a form of unsupervised machine learning, to identify universal indicators of deformation. This approach enables the detection of the onset of tertiary creep phase (deformation acceleration towards final failure), prediction of failure time, and accurate prediction of the failure location on the material surface just before the tertiary creep phase begins. Among the techniques used—Principal Component Analysis (PCA), Independent Component Analysis (ICA), Factor Analysis (FA), Non-negative Matrix Factorization (NMF), and Dictionary Learning (DL)—PCA and FA perform better in both detecting the onset of tertiary creep and predicting failure locations. The comparative analysis reveals the presence of universal characteristics in the evolution of local strain fields, offering a novel framework for studying material mechanics and providing key insights into failure prediction. In particular, the prediction of failure location as well as the comparison of the efficacy of various dimensionality reduction techniques are clearly novel aspects introduced in this work. • Universal characteristics of local strain field evolution were revealed. • Failure time and location were predicted through dimensionality reduction analysis. • Failure location was predicted well before visible macroscopic localization. • Principal Component Analysis and Factor Analysis outperformed other methods. • The approach employing DIC strain field data is applicable also to other materials.