Microstructure Identification of Additive Manufactured Titanium Alloy by Using Lamb Wave-DenseNet Network

TL;DR

This paper introduces a new method using ultrasonic testing and deep learning to accurately identify microstructures in 3D-printed titanium alloys, improving their mechanical reliability.

Contribution

A novel Lamb wave-DenseNet network is proposed for intelligent microstructure identification in additive manufactured titanium alloys.

Findings

The enhanced Lamb wave-DenseNet network achieved 97.93% accuracy in identifying three grain microstructure categories.

Simulation and experimental validation confirmed the method's effectiveness for large-scale engineering applications.

Abstract

In the additive manufacturing (AM) process, dynamic fluctuations in process parameters often result in non-uniform grain sizes in the microstructures of fabricated components, which impairs their stability of mechanical performance. Consequently, the accurate identification of microstructures in AM titanium alloy components is essential for optimizing their mechanical reliability and prolonging their service life in engineering applications. An approach combining ultrasonic testing and deep learning is provided to address the demands for high efficiency and intelligent identification of diverse grain microstructures in AM titanium alloys. First, the Centroidal Voronoi Tessellations (CVT) algorithm was employed to construct three representative simulation models that replicate the characteristic grain microstructures of AM titanium alloys encompassing fine-grained, coarse-grained, and mixed-grained configurations. Subsequently, COMSOL Multiphysics software (v.6.3) was utilized to perform laser-induced ultrasonic Lamb wave (LIULW) testing simulations on the CVT-based microstructure models. Further, a comprehensive simulation dataset was established, including time-domain signals and their frequency-domain features of LIULW. This simulation dataset was then used to train a neural network with an improved architecture, aiming to enhance the discriminative capability for subtle differences in LIULW signals induced by varying grain sizes. Experimental validation results demonstrated that the proposed enhanced Lamb wave-DenseNet network achieved an overall recognition accuracy of 97.93% for the three distinct grain microstructure categories. Collectively, these findings confirm that the integrated method provides a robust theoretical framework and a practical technical solution for large-scale, engineering-level microstructure identification of AM titanium alloy components. This work not only bridges the gap between microstructural simulation and intelligent LIULW testing but also lays a foundation for quality control in high-volume AM of titanium alloy structural parts.

Figures