AI-assisted NDE solutions have taken a firm standing in the NDE world. Currently, the industry has a lot variance in the utilization of these solutions. Some have years of experience working with modern ML-based analysis tools, while others are still considering getting started. While it has been shown that the technology is capable of producing real benefits, many AI projects fail in practice.

This presentation analyzes learnings over several years of AI/ML deployments across multiple industries and NDE modalities. These industry examples are used to collect a recipe for success including target selection, problem framing, human-machine interface, regulatory issues and more. Additionally, a collection of common pitfalls is presented with strategies for how to avoid them.

Nondestructive testing (NDT) is the essential, silent guardian of natural gas infrastructure, providing the only way to ensure pipeline integrity without interruption. Of the various NDT methods, X-ray (radiographic) weld inspection is the most common and foundational to pipeline safety. Human factors affect the interpretation of radiographs and can vary under this demanding perceptual-cognitive task that depends on the interpreter’s knowledge, mental state, visual perception, and the environment in which they work. NYSEARCH/NGA, in collaboration with Ooga Technologies, is conducting a structured validation of AI-based assisted weld defect recognition for natural gas pipeline radiography. The study establishes a human-reviewed ground truth from a curated set of digitized images and then evaluates AI performance against that baseline to quantify accuracy, sensitivity, and repeatability. We will share the test design, dataset curation approach, operator interpretation protocol, and evaluation metrics, along with lessons learned. The intended outcome is to demonstrate how Local Distribution Companies (LDCs) can integrate this technology into their inspection processes to enhance safety, improve efficiency, and prepare for a future where AI can reliably support human decision-making. This work is among the first structured efforts in North America to benchmark AI weld defect recognition against operator-reviewed radiographs within the natural gas sector.

Laser-based ultrasonic inspection offers a non-contact and full-field means of quantitatively evaluating guided-wave propagation in structures with complex geometries. The Quantitative Laser Ultrasound Visualization (QLUV) system developed at the National Taipei University of Technology integrates pulsed-laser excitation with ultrasonic detection to reconstruct high-resolution, time-resolved wavefields. Grounded in the reciprocity principle, QLUV visualizes Lamb, Rayleigh, and wedge wave propagation in metallic and composite structures. By capturing both amplitude and phase distributions of the wavefield, QLUV enables quantitative reconstruction of scattering behaviors and defect-induced wave interactions. It has demonstrated quantitative 2D and 3D imaging of scattering and reflection from machined holes and delaminations, providing a robust foundation for data-driven interpretation.

To achieve automated and objective defect evaluation, deep learning was incorporated into the QLUV workflow. The networks were trained on high-resolution, time-resolved wavefield data acquired from the QLUV system. A pixel-level convolutional segmentation network localizes and delineates defect regions, while a masked-region regression network estimates defect depth, defined as the distance below the surface for subsurface corrosion or drilled defects. The framework was trained and experimentally validated through specimen-wise cross-validation on aluminum and stainless-steel plates containing drilled and corrosion defects, ensuring model generalization across materials and defect types. Results indicate that the deep-learning–enhanced QLUV system reduces analysis time by approximately 90 % and decreases sizing error by 35–50 %, achieving a mean average precision (mAP@0.5–0.95) of 0.783 and a mean squared error (MSE) below 0.006.

This integration transforms QLUV from qualitative visualization into a deep-learning–driven quantitative framework for defect characterization and material evaluation. It enables automated recognition, quantitative sizing, and operator-independent assessment, demonstrating readiness for deployment in smart-manufacturing nondestructive testing workflow.

Despite the transformative potential of machine learning (ML) in ultrasonic non-destructive testing (UNDT), its deployment in this safety-critical industrial remains limited by its “Blackbox” nature. The inherent complexity of modern ML architectures obscures the interpretability of their predictions for NDT engineers, limiting the establishment of a clear inspection qualification methodology for AI-driven testing decisions. Furthermore, the lack of traceability in these models introduces uncertainty regarding their behavior on unseen data and raises fundamental concerns regarding their robustness.

To bridge the gap between predictive performance and inspection qualification, this presentation outlines a pathway toward interpretable, explainability-integrated ML in NDT frameworks. Building on recent work in explainability-regularized training for corrosion-profiling models, and ongoing developments in data-driven artefact suppression, this presentation demonstrates: 1) how explainable AI (XAI) techniques can be operationalized to quantify and enhance ML robustness against corrupted signal features, 2) how ML models can be regularized to focus on physically meaningful A-scan echo features, thereby reducing reliance on spurious artefacts during inference.
This presentation also discusses future directions for advancing explainability-integrated ultrasonic NDT based on the discussed progress. As the existing framework employs a manually defined corruption window around the backwall echo, the next step would be investigating automated identification of salient signal features without prior assumptions. We are also aiming to extend the explainability-regularized loss to more complex ML architectures. At the meantime, ongoing work is exploring a flexible toolkit for restoring signals affected by ultrasonic artefacts, rather than simply removing them. This toolkit will incorporate physical features to guide the model’s explanations and will be designed to remain robust when processing uncertain or noisy data collected in the field.