Machine learning is transforming manufacturing metrology by enabling data-driven modeling, automation, and real-time decision-making across the measurement process. This keynote reviews recent advances and future directions for integrating machine learning (ML) throughout the measurement workflow—from system setup to decision-making—by structuring the analysis of the state of the art using a data flow framework. Key applications include ML-assisted setup and calibration, enhanced measurement, virtual measurements, and classification-based inspection. The remaining key challenge is the integration of metrological traceability, standardized uncertainty quantification, explainability, and reproducibility. Bridging the underlying conceptual gap in understanding and evaluating uncertainty is essential to establish scientifically rigorous and industrially reliable ML-driven metrology for future manufacturing systems.

Various regularization terms have been proposed for inverse problems. In this work, we investigate the shearlet-based regularization in the context of sparse-view X-ray computed tomography(XCT) reconstruction of industrial parts. The shearlet transform is a multiscale and multidirectional representation that promotes sparsity in anisotropic features such as edges and textures. Consequently, the shearlet-based regularization effectively preserves fine details (e.g., pores) and texture information while suppressing under-sampling artifacts. The resulting regularized optimization problem is solved using the primal-dual algorithm Condat-Vũ, a very general algorithm that enables the incorporation of various regularization terms. While classical XCT reconstruction methods such as FDK tend to become highly unstable under sparse-view conditions, the proposed shearlet-based method outperforms the baseline method FDK, yielding improved reconstruction image quality and reduced streak artifacts. Fine details are preserved and under-sampling artifacts are mitigated. Moreover, when the number of X-ray projections is reduced to less than 10% of the full industrial acquisition, our results demonstrate that acceptable reconstruction quality can be achieved.

This paper introduces a Dual-Energy Swin U-Network architecture (DE-Swin U-Net), applied to Dual-Energy Computed Tomography (DECT) scans, to mitigate image artefacts generated during the tomographic inspection of multi-material assemblies or components. Recently, vision transformers have been replacing more traditional convolutional neural network architectures in the field of computer vision, due to their ability to model long-range relationships and global context via self-attention. Considering this, the authors propose a model architecture consisting of an encoder-decoder framework: the encoder, based on a Swin Transformer, extracts features and patterns from the collected dual-energy 2D projections, while the decoder reconstructs the final image from these features. The model is trained via a custom data generation pipeline, where simulated CT projections, at two energy levels, are generated from CAD models of electrical components. In addition, an ideal monochromatic reconstruction is also produced, serving as the ground truth. The proposed method demonstrates preliminary reconstruction performance of DECT scans, particularly with respect to reconstruction speed and the structural similarity index measure.

An essential step to maximize value recovery in Electrical and Electronic Equipment (EEE) is grading it based on surface defects for reuse. Although state-of-the-art computer vision-based methods can overcome manual inspections' subjectivity, they are limited by different scales of defects and time-consuming segmentation annotation of small defects. This study addresses these challenges by proposing a Surface Condition Evaluation System (SCES). SCES refines manual grading guides for quantitative evaluation, enhances YOLOv8 with proposed Defect Channel and Spatial (DCS) attention modules and Slicing Aided Hyper Inference (SAHI) to improve defect detection, then cascades Segment Anything Model 2 (SAM2) with Zhang-Suen thinning algorithm to segment and measure small defects without segmentation annotations. SCES achieves the detection average precision (AP) of 98.9 %, 81.2 %, and 80.5 % for missing_battery, missing_cover, and scratch on the collected laptop surface defect dataset, respectively, which are higher than other state-of-the-art methods. Additionally, the measurement accuracy for scratches reaches over 85 %, verifying the reliability of the proposed method. The results indicate that SCES employs novel computer vision methods to realize the automated and quantified product surface condition grading. This research supports the transition from manual evaluation to an efficient and objective automated approach in the EEE reuse process, thereby contributing to resource conservation.

Industries are increasingly adopting metallic laser powder bed fusion (PBF-LB/M) to fabricate complex components not achievable via traditional manufacturing. However, enhancing productivity by increasing laser energy density can lead to keyhole formation, causing pore defects from unstable vapor depressions in the melt pool. This study presents a data-driven surrogate model by correlating in-situ melt pool monitoring images with keyhole pores identified via X-ray computed tomography (X-CT). A significant challenge addressed is the spatialtemporal misalignment between in-process monitoring signals and final pore locations, as gas bubbles can migrate within the liquid melt pool before solidification. To mitigate this, a many-to-one approach is introduced, linking multiple melt pool frames to each defect location to incorporate the full thermal history relevant to defect formation. Twelve scan tracks were fabricated atop a Ti6Al4V cuboid, inducing keyhole pores by halving the scan speed or doubling the laser power, shifting from conduction mode to unstable keyholing mode. Experimental results demonstrate that a Random Forest model (Type I) utilizing physics-informed melt pool features outperforms a deep learning-based ResNet-LSTM model (Type II), achieving superior predictive accuracy (AUROC = 0.95, AUPRC = 0.92) and requiring notably less computational resources (7.5 times faster training and 27.6 times faster prediction). The key findings emphasize that leveraging physics-informed features and thermal history effects not only enhances prediction accuracy but also provides interpretability and computational efficiency, making this approach particularly suitable for future in-situ defect detection and qualify-as-youbuild process control in PBF-LB/M.

Surface defects in metallic laser powder bed fusion, including dross, staircase, and edge effects, undermine internal channel integrity, highlighting the need for as-built qualification to enable correction. Existing multiphysics simulations and trial-and-error experiments are time-intensive and inflexible for process control. This study targets actual-to-nominal errors within dimensional tolerance, proposing a high-fidelity surrogate model to predict deviations using melt pool monitoring data. Process signatures, including shape and thermal-related gray value features, were mapped onto CAD geometries. To enable localised prediction, feature volumes were divided into cuboids spanning multiple scan tracks and layers, capturing inter-hatch and inter-layer interactions. Deviation labels quantified via X-ray computed tomography allowed supervised training. The surrogate integrates a convolutional neural network (CNN) to capture spatial dependencies across scan tracks and a Transformer encoder to model inter-layer remelting. This configuration achieves root mean squared errors of 6.1 mu m (11.3% of the 54.1 mu m average ground-truth deviation) for up-skin and 27.4 mu m (9.4% of 291.7 mu m) for down-skin surfaces, with R2 of 0.91 and 0.95, respectively. Ablation studies confirm CNN's intra-layer and Transformer's inter-layer contributions. Visualisation techniques (Grad-CAM and attention maps) highlight critical regions and layers, enhancing model interpretability and demonstrating its potential for in-situ surface quality control.

Computed Tomography (CT) reconstruction is an important inverse problem in industrial imaging, requiring robust methods to address different sources of error in the data and model. Among the various reconstruction approaches that tackle different challenges in CT modeling [1], such as limitations in the data, statistical methods are known for their ability to model various properties [2, 3] of the input data and assess their impact on the reconstruction outcome [4]. These methods can also incorporate regularization techniques, which, for instance, force smoothness in the solution by penalizing heavy oscillations in pixel values. Given our interest in the uncertainty Quantification (UQ) of CT reconstruction, we employ statistical methods within the Bayesian framework for solving this inverse problem. First, we define a model for CT reconstruction and formulate a statistical framework for the problem. We then reconstruct our data using appropriate methods and quantify the uncertainty in the results. Our primary focus is on the effect of noise on the reconstruction and the corresponding uncertainty it induces. Additionally, we aim to utilize computationally feasible sampling techniques to analyze the distribution of the solution, enabling an in-depth evaluation of the results. To achieve these goals, we apply a rapid regularized Markov Chain Monte Carlo (MCMC) reconstruction method [4, 5], employing the Metropolis-Adjusted Langevin Algorithm (MALA) [6] and its Lipschitz-adaptive variant (LipMALA) [7]. This approach produces a volumetric model where each voxel is represented by a probability distribution, which can be transformed into a triplet of gray-value models: one for the central value and two for the bounds of the confidence interval. Bi-directional and uni-directional length measurements derived from these gray-value models applied to real CT data yield task-specific measurement uncertainties. This method significantly reduces computational and storage demands compared to classic Monte Carlo simulations while incorporating regularization techniques. Experimental results using aluminum cylindrical step gauge [8] data acquired with a Nikon 225 CT system validate our approach. This work provides a scalable, statistically grounded methodology for UQ in CT reconstruction, offering enhanced reliability for industrial applications.

This work examines the impact of powder particle size distribution (PSD) on the surface finish of maraging 300 steel parts produced by laser powder bed fusion (LPBF). After an initial selection of two suitable powder raw materials, Mar 10-30 and Mar 15-45, the study demonstrates that finer powders lead to significantly smoother upfacing surfaces, with up to a 50 % reduction in surface roughness (Ra). The Mar 10-30 distribution not only produces superior surface quality but also allows for more advanced LPBF processing strategies. These strategies combine small layer thicknesses with laser remelting, demonstrating that top surface roughness can be reduced to as low as 1.5 mu m Ra. Finally, an outlook is given, suggesting that properly engineered powders, together with optimized LPBF processing parameters, enable the fabrication of both vertical single-track ridges or thin-walls down to similar to 100 mu m thickness and basic texturing patterns, offering a cost-effective insitu LPBF alternative to post-processing texturing. These developments open new possibilities for the surface finish and functionalization of LPBF components while increasing the precision and feature resolution in LPBF parts.

Modern manufacturing processes are monitored by different types of sensors throughout the entire process chain eventually sensing complementary pieces of information which results in the necessity to process multimodal sensor data. Specifically in the context of additive manufacturing, a combination of in-situ sensors and ex-situ measurement systems are used to capture both external contours and internal structures of the manufactured component. As a result of using heterogeneous measurement principles, the acquired sensor data is diverse and multimodal. Typically, this multimodal sensor data is analyzed independently from each other focusing on different quality characteristics. However, it is essential to consolidate this multimodal sensor data into a unified data model to achieve holistic quality assurance. This allows for evaluation of component quality at any stage in the manufacturing process, thereby enabling a more holistic approach to quality assurance. The aim of this study is to integrate multimodal sensor data into a newly developed consistent voxel-based data model. A holistic quality assurance for Fused Deposition Modeling can be realized in-situ by using a 2D camera and a laser light section sensor and ex-situ using X-ray computed tomography. The distinct datasets are then aligned and merged into a unified data model that incorporates both nominal and sensor-derived information. This unified voxel-based data model can serve in further investigations as the foundation for voxel-specific evaluation of the quality and the application of AI-driven quality analysis techniques.

X-ray computed tomography (CT) has gained significant traction across various industrial sectors as a dimensional metrology tool, providing holistic characterizations and enhancing the development of new industrial products. However, despite its potential, dimensional CT scanning can be relatively slow, often hindering routine adoption for many applications. This study investigates strategies to optimize CT measurement speed while maintaining measurement accuracy through an experimental approach that includes fast CT scanning configurations, fast CT simulation-based technique, and the adoption of an iterative reconstruction method. Using an additively manufactured metal part as a case study, the research evaluates measurement deviations across different geometrical features. Results indicate promising performance of fast CT scanning for dimensional measurements, as well as the feasibility of using a digital representation to support a fast CT measurement process. Additionally, the constraint split Bregman iterative reconstruction method combined with total variation demonstrates significant improvements over traditional Feldkamp, Davis, Kress reconstruction, achieving up to 90% accuracy enhancement for external features with limited X-ray projections. These findings provide valuable insights for implementing efficient CT metrology in contemporary industry where both speed and accuracy are essential.

Statistical methods within the Bayesian framework have been widely used to address inverse imaging problems, such as computed tomography (CT) image reconstruction. These methods offer a probabilistic approach that is able to enhance the reconstruction quality by employing regularization methods while enabling uncertainty quantification of the result, providing valuable insights into the reliability of the reconstructed images. However, despite the flexibility and range of techniques within this framework, the computational intensity of this class of approaches is still impractical for large-scale datasets like those in CT. In this manuscript, we introduce a concept for determining the uncertainty caused by the noise in the observed data in CT-based dimensional measurement using a rapid, regularized, Markov Chain Monte Carlo reconstruction technique. This method provides a volumetric model where each voxel is represented by a distribution, which is then transformed into a triplet of gray value models: one for the central value and one each for the upper and lower bounds of the confidence interval. Bi-directional and uni-directional length measurements on results derived from each single-gray-value model, for real CT data, provide a task-specific measurement uncertainty. This method requires significantly less computation and storage capacity compared to classic Monte Carlo simulations by reducing the number of needed simulations for approximating a distribution while incorporating regularization techniques. The results are compared to conventional non-regularized and regularized reconstruction methods, such as Feldkamp–David–Kress (FDK), and state-of-the-art statistical methods, followed by validation of the determined uncertainty in real CT data.

In every product reuse, defect detection and analysis of the severity of the detected defects are essential to correctly assess the product’s grade, which for most high-end electronics also defines to a large extent their residual value. However, the detection and measurement of small or narrow defects, such as scratches, remains a challenge for commonly adopted deep-learning segmentation approaches. In addition, pixel-level annotation that is required for the training of neural networks is highly time-consuming and labor-intensive. Therefore, to mitigate the subjectivity of purely manual defect inspection using human eyes and to reduce the labeling time required for training neural networks to perform this task, a method is proposed that combines an object detector with a Large Vision Model (LVM). This method allows the training of model with only bounding box annotations, while still producing segmentation masks that contour the scratches with high accuracies as output. Furthermore, for measuring scratch lengths, the K3M thinning method is utilized to compute the arc length of each detected scratch. To validate the effectiveness of the proposed method, a laptop dataset containing 220 images with 2,971 scratches is collected. Results demonstrate that the cascaded method adopting Faster R-CNN-FPN and the Segment Anything Model (SAM) achieves an accuracy of 83.7% for the scratch length measurement, which corresponds to a mean absolute error of 1.60mm with a measurement accuracy of 91.5%. These results indicate the great potential of the use of an LVM-driven approach for defect detection and enhancing the grading accuracy and objectivity in electronics reuse.

To extend the applicability of in-line computed tomography (CT) within Industry 4.0, accelerating the data acquisition and image reconstruction process is essential to meet the demands of real-time, high-throughput inspection. This paper focuses on accelerating in-line CT by addressing the trade-off between image quality and angular sampling reduction through the development of dedicated reconstruction algorithms. Various inherent properties of in-line CT are leveraged as a priori knowledge, specifically the noise level, within a total variation (TV)-based reconstruction framework to enhance reconstruction quality, support automation, and enable accurate image analysis using 2%-5% of the data required by standard methods such as Feldkamp-Davis-Kress (FDK).

X-ray computed tomography (XCT) is a validated and frequently used tool to verify part geometry and to perform a non-destructive inspection of additive manufacturing parts. However, the acquisition of a large number of x-ray projections generates long inspection times. This conflicts with a high throughput of the production process and hinders the integration of XCT as an in-line quality control procedure for low-end parts. In this paper, we propose a method to obtain three-dimensional (3D) information of the internal pores from a limited view or limited angle scan. The method combines a forward projection model of a cone-beam x-ray system and a deep learning neural network to directly classify each individual voxel, based on the x-ray projections in order to avoid the reconstruction and segmentation step. Accompanying reconstruction artifacts for limited view and limited angle XCT scans are thereby reduced, while preserving 3D information of the pores, defects or inclusions inside the material. The method is validated on real x-ray projections of polymer laser sintered industrial parts and shows a significant reduction in the required x-ray projections, hence acquisition time.

Laser powder bed fusion, while promising, faces hurdles in certifying fabricated parts due to cost and complexity, with in-process monitoring emerging as a potential solution. Existing models focus on predicting defects at a given location using the monitoring signals from solely that same location. Hence, these models treat each track or layer independently of the previous and subsequent ones, neglecting potential interdependencies. This study proposed an in-situ, photodiode-based monitoring approach considering inter-hatch and inter-layer effects on porosity formation- factors often overlooked in existing research. Two Ti-6Al-4 V cuboids (10x10x5 x10 x5 mm3) 3 ) were built with optimized process parameters, with the melt pool continuously monitored at 20 kHz via a co-axially mounted photodiode. The monitoring system captured the integral radiation in the near-infrared spectrum within a field of view centered on the melt pool. The porosity is assessed by X-ray computed tomography (X-CT), serving as ground truth to build supervised machine learning (ML) models. This study considered physical phenomena occurring during the printing process, including remelting of lack of fusion pores by the subsequent layer, keyholes penetrating the current layer hence introducing pores in the layer below, and overlap between adjacent scan tracks. These considerations are critical for a holistic understanding of pore formation mechanisms. Photodiode signals and computed tomography volumes were cropped using windows of four sizes to test the model's pore localization capability. A machine learning model, specifically a Convolutional Neural Network (CNN)- Long Short-Term Memory (LSTM) network, was trained to predict porosities using these window sequences. The CNN extracted spatial features from photodiode signals, addressing inter-hatch effects, while the LSTM captured temporal dependencies across layers, addressing inter-layer effects. The results, with the Area Under the Receiver Operating Characteristic curve (AUC) of 0.91 for pores exceeding 8000 mu m3 3 in volume and 100 mu m2 2 in cross-sectional area, demonstrate the feasibility of the proposed model in detecting pores-level defects. This high defect prediction and positioning accuracy are essential for process control, providing real-time status of the region of interest and informing the controller of pore positions, thus facilitating intra-layer or inter- layer correction.

This work is concerned with fan- and cone-beam computed tomography with circular source trajectory, where the reconstruction inverse problem requires an accurate knowledge of source, detector and rotational axis relative positions and orientations. We address this additional inverse problem as a preceding step of the reconstruction process directly from the acquired projections. In the cone-beam case, we present a method that estimates both the detector shift (orthogonal to both focal and rotational axes) and the in-plane detector rotation (over the focal axis) based on the variable projection optimization approach. In addition and for the fan-beam case, two new strategies with low computational cost are presented to estimate the detector shift based on a fan-beam symmetry condition. The methods are validated with simulated and experimental industrial tomographic data with code examples available for both fan- and cone-beam geometries.

By detecting internal batteries before treatment, the end-of-life handling of Waste Electrical and Electronic Equipment (WEEE) can be made more safe and cost efficient. Where prior research has demonstrated the potential of deep learning models to detect batteries in X-ray transmission (XRT) images of WEEE, these studies did not fully investigate the potential of the higher bit depth that XRT machines are capable of. A 16 bit bit-depth enables the representation of 65,536 shades in greyscale, unlike the 8 bit bit-depth that is commonly used in deep learning approaches, which is limited to 256 shades. Therefore, the presented research investigates the impact of higher bit-depth on the performance of battery detection in WEEE. In addition, the detection and classification performance of batteries is tested across two sets of X-ray source parameters (120kV and 100mA; 60kV and 40mA). Results show that, for this use case with the same set of source parameters, using 16 bit image data yields small but consistent improvement in overall mAP50 and mAP50-95 as compared to using 8 bit image data. Hence, the findings of the presented research contribute to advancing WEEE recycling processes through improved battery detection in XRT images, contributing to more efficient resource recovery.