Porosity evaluation of additive manufacturing components by deep learning-assisted nonlinear ultrasonic technique

Abstract The porosity of additive manufacturing (AM) components exerts a profound influence on their mechanical properties, thereby impeding their widespread engineering adoption. Current porosity evaluation methodologies predominantly rely on destructive testing approaches, underscoring the pressing demand for a non-destructive testing (NDT) method enabling precise and in-situ porosity evaluation. In this paper, we propose a novel method for porosity evaluation of AM components by deep learning (DL)-assisted nonlinear ultrasonic technique. Experimental investigations revealed that ultrasonic nonlinear responses exhibit pronounced sensitivity to subtle variations in AM component porosity. However, the complex mapping between porosity and the nonlinear response of ultrasonic waves, stemming from the intricate interplay of multiple printing parameters, poses a significant challenge for traditional modeling approaches relying solely on the nonlinear coefficient. To address this challenge and enable high-precision porosity evaluation, we established a Convolutional Neural Network (CNN) - Bidirectional Long Short-Term Memory (BiLSTM) - Attention network. The nonlinear response of ultrasonic waves was extracted using the phase-reversal method and employed as the proposed network's input. The trained network achieved high-precision prediction of porosity. Comparative analyses between the predicted results obtained by using phase-reversed signals and conventional evaluation signals highlighted the indispensable role of nonlinear ultrasonic responses in capturing porosity-dependent features, providing a explanation for the established DL method. The proposed method reveals the complementary integration of nonlinear ultrasonic waves and the DL method: the former provides physical sensitivity to porosity evolution, while the latter resolves the ill-posed nature of inverse problems. This synergistic framework establishes a novel solution for high-precision NDT of AM components.