Erratum to: Comparative analysis of deep learning models for defect detection in additive manufacturing using thermal imaging (Results in Engineering, (2025), 28, C, (108359), (S2590123025044032), 10.1016/j.rineng.2025.108359)

The adoption of additive manufacturing (AM) in critical applications is hindered by quality control challenges, where defect detection systems must balance accuracy, computational efficiency, and real-time deployment requirements [1]. Despite numerous individual convolutional neural network (CNN) implementations for thermal defect detection, the lack of systematic architectural comparisons creates uncertainty about optimal model selection for different manufacturing contexts. In extrusion-based AM techniques such as Direct Ink Writing (DIW), defects like over-extrusion (excess material causing beads and uneven surfaces) and under-extrusion (insufficient material leading to inter-layer gaps) frequently occur, degrading mechanical properties and surface finish [1]. These defects not only weaken fabricated parts but also result in wasted material and time if undetected [1]. DIW enables the processing of paste-like materials, including ceramics, polymer composites, and functional inks, facing thermal monitoring challenges related to material flow, layer adhesion, and dimensional accuracy [2]. Hence, early detection, ideally during printing, is vital to enable corrective actions before irreversible damage occurs. Consequently, there is a growing focus on the development of in-situ (real-time) monitoring and defect detection strategies to enhance the reliability of additive manufacturing processes. Among in-situ sensing modalities, such as optical, acoustic, and profilometric systems, infrared (IR) thermal imaging has emerged as a particularly effective approach for detecting defects during 3D printing [1,3]. IR thermal imaging provides several technical advantages for real-time defect detection: it enables non-contact, continuous monitoring of temperature distributions across the build surface without interfering with the printing process; it captures thermophysical phenomena directly related to defect formation, such as abnormal heat accumulation and dissipation patterns; and it operates effectively in enclosed printing environments where optical access may be limited [4,5]. IR cameras capture real-time temperature distributions, and thermal signatures often correlate with print quality: poor layer adhesion or internal voids manifest as atypical cooling patterns or localized hotspots. AbouelNour et al. [3,6] reported that printed polymer specimens with engineered internal defects exhibited higher surface temperatures and more pronounced hotspots compared to defect-free samples. These findings indicate that IR thermography can serve as a reliable early warning method for defect formation. The integration of machine learning (ML), including Deep Learning (DL), with thermal monitoring has produced numerous ML-based defect detection approaches, yet the field lacks systematic architectural benchmarking to guide implementation decisions [7]. Recent advances in deep learning for AM quality control have demonstrated significant progress across multiple domains. Bhandarkar et al. [4] developed CNN-based warpage detection systems achieving high accuracy in identifying dimensional deviations during polymer printing, while Bhandarkar et al. [5] provided comprehensive reviews of both traditional and advanced defect detection methodologies, highlighting the transition from rule-based to learning-based approaches. Bhandarkar et al. [8] conducted comparative investigations of deep learning models for defect detection in polymer parts, demonstrating the effectiveness of modern architectures. Bhandarkar et al. [9] implemented real-time monitoring frameworks that significantly reduce material wastage through early defect identification, and Singh and Desai [10] pioneered automated surface defect detection using integrated machine vision and CNN systems. However, existing studies typically evaluate single architectures in isolation, making it impossible to assess relative performance, computational trade-offs, or deployment suitability across different ML Approaches [11,12]. This fundamental gap in comparative evaluation motivates our systematic benchmarking study. Our approach employs a systematic comparison of seven distinct deep learning approaches for detecting five categories of defects that happen on AM, trained using the Similar framework with similar optimization techniques. The defect categories are chosen carefully from existing research by observing different defects [13,14]. The following five defect categories are considered in our benchmarking study, each representing a distinct challenge in additive manufacturing defect detection:(1)Under-extrusion: Occurs when the printer extrudes insufficient material, resulting in gaps, weak layers, or incomplete structures.(2)Over-extrusion: Happens when excess material is extruded, leading to blobs, rough surfaces, and dimensional inaccuracies.(3)Warping: The printed part curls or lifts at the edges due to uneven cooling, affecting dimensional precision.(4)Layer Shifting: Misalignment between successive layers caused by mechanical slips or sudden interruptions.(5)Normal Print: A defect-free print with uniform extrusion, proper layer alignment, and accurate dimensions. Under-extrusion: Occurs when the printer extrudes insufficient material, resulting in gaps, weak layers, or incomplete structures.