X-ray and neural network based in-situ identification of the melt pool during the additive manufacturing of a stainless steel part

Loïc Jegou, Valérie Kaftandjian, Thomas Elguedj, Mohamed Tahraoui, Philippe Duvauchelle, Mady Guillemot · 2025

Read X-ray and neural network based in-situ identification of the melt pool during the additive manufacturing of a stainless steel part by Loïc Jegou, Valérie Kaftandjian, Thomas Elguedj, Mohamed Tahraoui, Philippe Duvauchelle, Mady Guillemot (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-22409.txt.

Introduction

Loic Jegou and 5 co-authors watch a melt pool form inside metal as a part is printed. The process, laser metal deposition with powder, builds or repairs steel parts layer by layer. Pool shape shapes the final part's quality, yet it hides inside the metal. X-ray imaging can see through the part and show the pool by its density contrast.

The team filmed ten layers of stainless steel as they were laid down. The X-ray beam's mixed energies blurred contrast, so classic image methods failed. Neural nets trained on simulated heat and X-ray data found the pool instead. A VGG16-style net worked well, backing live X-ray checks in factories.

Catalog record

CREATORLoïc Jegou, Valérie Kaftandjian, Thomas Elguedj, Mohamed Tahraoui, Philippe Duvauchelle, Mady Guillemot
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-22409

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

X-ray and neural network based in-situ identification of the melt pool during the additive manufacturing of a stainless steel part Loïc Jegou, Valérie Kaftandjian, Thomas Elguedj, Mohamed Tahraoui, Philippe Duvauchelle, Mady Guillemot Abstract Laser Metal Deposition with Powder (LMDp) is an additive manufacturing technique used for repairing metal components or producing parts with intricate geometries. However, a comprehensive understanding of the melt pool dynamics, which significantly influences the final properties of LMDp-fabricated parts, remains limited. Non-destructive testing is highly valuable for conducting in-situ controls during manufacturing. X-ray imaging offers the ability to penetrate metallic parts and detect defects such as porosity. In the context of additive manufacturing, X-rays can be employed to visualize the shape of the melt pool during the fabrication process. The contrast between the liquid and solid phases, due to their density differences, should be observable in the radioscopy images. The experimental setup required to perform such a test on an industrial additive manufacturing installation consists of a movable X-ray source that produces polychromatic beams, a detector, and extensive lead shielding to ensure X-ray safety. In-situ observations of the melt pool were conducted during the deposition of ten successive layers of stainless steel 316L (SS316L). The polychromatic nature of the X-ray beam, however, rendered traditional image analysis methods ineffective for detecting contrast variations. To address this challenge, neural networks trained on simulated data (thermal and X-ray) were employed, providing a solution to identify the melt pool in low-contrast radioscopic images. The architecture inspired by VGG16 demonstrated promising results, confirming the potential for in-situ non-destructive testing using X-ray imaging in industrial additive manufacturing processes. keywords Additive manufacturing, Direct Energy Deposition, Stainless steel, X-ray, Melt pool, Convolutional neural networks 1 Introduction Additive manufacturing (AM) is defined by the American Society for Testing and Materials as: ”the process of joining materials to make objects from 3D model data, usually layer upon layer, as opposed to subtractive manufacturing methodologies, such as traditional machining”. While this definition applies to all types of materials, this work specifically focuses on metallic parts. Regarded as the new industrial revolution Berman 2012 , AM has continuously evolved to address current industrial demands, such as rapid production, material efficiency, and the ability to create complex, custom-designed components. Today, a lot of different technologies exist Frazier 2014 , and this article focuses on Laser Metal Deposition with Powder Ahn 2021 (LMDp). LMDp involves using a highly focused laser to melt a localized region of a substrate, while continuously delivering powdered feedstock material into the melt pool Svetlizky et al. 2021 . During this process, there is a complex interaction between the powder, laser, and melt pool which is strongly influenced by the operating parameters (such as powder flowrate, laser power and scanning speed). These parameters significantly affect the quality of the final part in terms of geometry, residual stress Lu et al. 2019 , and metallurgical properties Marya et al. 2018 . The melt pool serves as a reliable indicator of defect formation Scime and Beuth 2019 and overall quality. Monitoring it in real-time during the manufacturing process is highly valuable, both for predicting defect locations (such as porosity, as demonstrated by Khanzadeh et al. 2019 using infrared images) and for adjusting operating parameters during the fabrication, as seen in studies using infrared sensors Gibson et al. 2020 or CMOS cameras Akbari and Kovacevic 2019 . Numerous imaging techniques have been developed to monitor the melt pool’s morphology Sun et al. 2020 ; Ding et al. 2016 ; Smurov et al. 2013 , its temperature Liu et al. 2014 ; Everton et al. 2016 , or both simultaneously Jegou et al. 2023 . While these methods provide critical insights into the melt pool’s behavior, they are limited to surface observations, leaving the subsurface dynamics unobserved. This is where X-rays step in, offering the ability to peer through the material and capture details beneath the surface. In the industry, X-rays are extensively used to perform Non-Destructive Testing (NDT) Hanke et al. 2008 to inspect manufactured parts for defects, such as porosity. Radioscopy is an imaging technique that creates two-dimensional grayscale images that reflect the degree to which X-rays interact with an object Letang 2023 . This method examines contrasts resulting from variations in thickness, chemical composition, or density. In metal AM, the idea is to analyze the melt pool by leveraging the contrast created by the density differences between the liquid and solid phases. Prior to its application in AM, X-rays were used to study melt pools in welding and laser cutting processes with similar objectives Abt et al. 2011 ; Boley et al. 2013 . While porosities were easily visible in these studies, the melt pool contours were often indistinct due to low contrast between the two phases. Yamada et al. 2012 addressed this issue using synchrotron radiation, where X-ray beams are far brighter (at least a million times more photons) than those produced by portable X-ray sources. This led to numerous studies of LMDp conducted at synchrotron facilities, such as the Advanced Photon Source in the USA Wolff et al. 2019 ; Wolff and Gould 2022 ; Lindenmeyer et al. 2021 ; Webster et al. 2023 and Diamond Light Source in England Chen et al. 2021a ; Chen et al. 2021b ; Chen et al. 2021c . These studies provided valuable insights into porosity formation mechanisms, powder particle interactions above the melt pool, particle trajectories influenced by laser-induced pressure gradients, and melt pool dynamics. However, synchrotrons are huge installations that generate intense X-ray beams with the capability to filter them into monochromatic rays. The downside is that AM setups must be miniaturized to fit within these facilities Dass et al. 2022 , which results in systems with characteristics that differ from those used in industrial environments. The primary goal of this study is to make X-ray NDT feasible on an industrial additive manufacturing setup, using a movable X-ray source. Portable X-ray sources, unlike synchrotrons, produce polychromatic beams that are significantly less bright, resulting in radioscopic images with lower contrast. The next challenge is to develop a method for identifying the melt pool in these low-contrasted grayscale images. This issue closely parallels challenges in the medical field, where tumors or pathological features must be detected in X-ray scans with similarly low contrast. Several image segmentation techniques exist, including thresholding Sahoo et al. 1988 , region-based methods Karthick et al. 2014 , boundary-based approaches Zhu et al. 2007 , and hybrid techniques Nyma et al. 2012 . Currently, automatic segmentation methods are favored due to their lack of reliance on manual adjustments. Otsu’s thresholding, known for its efficiency, has inspired many researchers Ng 2006 ; Fan and Lei 2012 . However, the effectiveness of these methods can be compromised by the poor quality of the radioscopy, the size of the tumor or the level of contrast Swetha and Bindu 2016 . Artificial Intelligence (AI) provides new approaches to deal with these segmentation challenges, achieving higher success rates Chowdhary et al. 2020 ; Liu et al. 2019 . Typically, computer vision problems are solved with Convolutional Neural Networks (CNN). A comprehensive review of these architectures is provided by Zhang et al. 2017 . Among these, auto-encoders (a specific type of

Read the complete text