# Deep Learning-Assisted Porosity Assessment for Additive Manufacturing Components Using Ultrasonic Coda Waves

## TL;DR

This paper introduces a non-destructive method using deep learning and ultrasonic coda waves to accurately assess porosity in 3D-printed components.

## Contribution

A novel deep learning-assisted non-destructive testing method combining ultrasonic coda waves and a multi-head attention mechanism for porosity assessment in additive manufacturing.

## Key findings

- Ultrasonic coda waves show high sensitivity to porosity changes in additive manufacturing components.
- The proposed coda-CNN-MAM network achieves 98% accuracy in porosity prediction.
- Traditional parameters fail to map porosity due to complex micro-structural features, but deep learning overcomes this limitation.

## Abstract

An innovative approach for porosity assessment of Additive Manufacturing (AM) components is proposed, integrating Deep Learning (DL) with ultrasonic coda waves.The integration of feature-rich ultrasonic coda waves with the strong feature extraction of DL enables highly accurate porosity evaluation for AM components.

An innovative approach for porosity assessment of Additive Manufacturing (AM) components is proposed, integrating Deep Learning (DL) with ultrasonic coda waves.

The integration of feature-rich ultrasonic coda waves with the strong feature extraction of DL enables highly accurate porosity evaluation for AM components.

What are the main findings?
Ultrasonic coda waves can propagate fully inside the AM components and exhibit high sensitivity to changes in the AM components’ porosity.The introduction of a Multi-head Attention Mechanism (MAM) effectively improves the porosity assessment accuracy of the proposed coda-CNN-MAM network.

Ultrasonic coda waves can propagate fully inside the AM components and exhibit high sensitivity to changes in the AM components’ porosity.

The introduction of a Multi-head Attention Mechanism (MAM) effectively improves the porosity assessment accuracy of the proposed coda-CNN-MAM network.

What are the implications of the main findings?
Nondestructive Evaluation (NDE) of porosity in AM components is achieved using ultrasonic coda waves.The integration of coda waves with DL provides a highly accurate framework for assessing porosity in AM components.

Nondestructive Evaluation (NDE) of porosity in AM components is achieved using ultrasonic coda waves.

The integration of coda waves with DL provides a highly accurate framework for assessing porosity in AM components.

The porosity of additive manufacturing components significantly impacts their mechanical properties, thereby limiting their widespread application in engineering. Current porosity assessment predominantly relies on destructive testing, underscoring the urgent need for accurate in situ non-destructive testing methods. In this paper, we propose a novel deep learning-assisted non-destructive testing method for porosity assessment in additive manufacturing components. Our approach leverages the high sensitivity of ultrasonic coda waves to minute internal material changes, combined with the powerful feature extraction capability of deep learning. Experimental results demonstrate that ultrasonic coda waves are sensitive to porosity variations in additive manufacturing components. Due to the porosity of additive manufacturing components involves multi-dimensional micro-structural features, conventional parameters such as the correlation coefficient and relative velocity change cannot establish an effective mapping relationship, despite their variation with porosity, thus precluding accurate inversion. To address this challenge, we propose a coda–convolutional neural network–multi-head attention mechanism network. Ultrasonic coda waves can fully interact with pores inside additive manufacturing components, and their signals are rich in porosity-related features. The introduction of deep learning significantly enhances the ability to extract such features. The trained network achieves high-precision porosity prediction with an accuracy of 98%. Our proposed approach reveals the complementary integration of ultrasonic coda waves and deep learning methods: the former provides high sensitivity to porosity changes, while the latter addresses the limitations of difficult extraction of relevant features and unclear complex mapping relationships. This collaborative framework establishes a new solution for high-precision non-destructive testing of additive manufacturing components.