Deep learning-enabled accurate and interpretable stress–strain characterization from ultrasonic measurements

초록

While ultrasonic measurements have been employed to analyze the material properties in a non-destructive manner, challenges remain in accurately predicting the mechanical properties of materials and understanding how diverse ultrasonic characteristics contribute to the predictions. This paper proposes a deep learning approach for accurate and interpretable characterization of the mechanical behavior of materials based on non-destructive ultrasonic measurements. Specifically, the proposed framework is designed to predict the stress–strain behavior of materials from multiple ultrasonic signals and to provide interpretability between non-destructive and destructive modalities. A multi-branch aggregated one-dimensional convolutional neural network is proposed to effectively capture signal-wise dependencies from a collection of distinct ultrasonic signals. Both quantitative and qualitative results demonstrate that the proposed model achieves superior predictive performance, showing improvements of up to 78% in mean absolute percentage error and 95.4% in R2 score compared to baseline models. Moreover, we introduce a signal-level gradient-retrieved interpretation method to reveal the contribution of each ultrasonic characteristics to the predicted stress–strain behavior. The proposed approach not only enhances predictive performance but also enables physically relevant interpretation, offering a promising method for accurate and explainable non-destructive evaluation of material properties.

키워드