Prediction of statistical force–displacement curves of Charpy-V impact tests based on unsupervised fracture surface machine learning

• Unsupervised machine learning fractography analysis corresponding to lower shelf, transition, and upper shelf regimes. • Extracted clusters strongly correlate with key parameters, enabling accurate classification of fracture behavior. • Explainable AI using a Grad-CAM analysis. • Statistical material-sensitive F-s curve inference with ± 1σ corridors. While conventional pendulum impact tests only measure a material’s integral energy absorption, the instrumented version of the test provides valuable additional insights by extracting force–displacement behaviour of the loaded specimen. The latter, however, requires auxiliary testing equipment, calibration procedures and evaluation methods. Therefore, this study aims to predict force–displacement behaviour of instrumented Charpy impact tests solely on the basis of analyzing images of specimen fracture surfaces postmortem. This is explored and achieved by using unsupervised machine learning techniques for computer vision. By using unsupervised computer vision on fracture images from 4 steels, we assess the feasibility of classifying fracture surfaces and deriving statistical force–displacement curves and provide crucial interpretability of the model’s decision making. The results indicate the model’s ability to learn the necessary representations without the need of supervision. The unsupervised model can extract significant insights from fracture images alone, supporting efficient, accurate, and interpretable materials testing, where confidence intervals of 97 % can already be met for the upper shelf. This allows detailed information about the mechanical behaviour of the material to be obtained from non-instrumented tests.