⚡ electrical engineering

Direct Geometric Reconstruction of Flaws from Simulated Ultrasonic Inspection Data

This study demonstrates a proof of concept for directly reconstructing the geometric contours of inner-surface-breaking flaws in steel from simulated pulse-echo ultrasonic data using a supervised convolutional regression model, bypassing traditional image reconstruction and interpretation.

Original authors: Daniel Algernon

Published 2026-10-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniel Algernon

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Modern infrastructure, from the pipelines carrying energy to the wings of aircraft, relies on materials that remain sound over decades of service. Before a crack becomes a failure, it must be found and measured, a task handled by nondestructive testing. Among the tools used for this, ultrasonic testing is a primary method. It works by sending high-frequency sound waves into a solid object. When these waves hit a flaw, such as a crack or a void, they bounce back to a sensor. The challenge for engineers is that the sound wave does not carry a picture of the crack; it carries a complex signal of echoes. To understand the size, depth, and angle of the damage, experts must interpret these echoes, often by building a visual image of the interior and then guessing the shape of the flaw based on that image. This process is indirect and relies heavily on human judgment.

A new study by Daniel Algernon from the Swiss Association for Technical Inspections asks a different question. Instead of building an image first and then interpreting it, can the sound signal be translated directly into the shape of the flaw? The researcher treated the problem as a direct geometric puzzle. Using powerful computer simulations, he generated thousands of examples where the exact shape of a crack was known, along with the corresponding sound signal it would produce. He then trained a specialized computer program to learn the relationship between the two. The goal was to see if the machine could look at the raw sound data and immediately sketch out the outline of the defect, bypassing the intermediate step of creating a visual image.

The study focused on a specific type of damage: cracks that start at the inner surface of a steel component and break through with a faceted, angular shape. The researcher used a standard inspection setup involving a single sound probe mounted on a wedge, which sends a beam of sound into the steel. In the simulation, the computer created over one thousand different crack shapes, varying their depth, tilt, and local profile. For each unique crack, the software calculated exactly what the sound signal would look like. This created a massive library of paired data: a sound signal on one side and the precise coordinates of the crack outline on the other. The computer program was then taught to find the pattern that links the signal to the shape.

The program used a type of artificial intelligence known as a convolutional regressor. Unlike systems designed to recognize cats or cars in photographs, this system was built to be compact and efficient. It took the sound signal, which was represented as a grid of data points, and output a list of eleven points that defined the crack's outline. The researchers did not ask the computer to guess the entire image of the crack; they asked it to estimate a specific set of coordinates that described the geometry. This approach is similar to asking a person to describe the shape of a shadow by listing the positions of its corners, rather than asking them to draw the whole object. The computer learned to make this list by studying the thousands of simulated examples, adjusting its internal settings until its guesses matched the known shapes as closely as possible.

When the researchers tested the trained program on new, unseen simulated cases, the results were promising but specific to the conditions of the experiment. The system successfully reconstructed the general shape of the cracks. It was particularly good at determining the tilt, or the angle, of the crack. The computer predicted the angle of the flaw with a high degree of consistency, often within a few degrees of the true value. This suggests that the overall orientation of the crack leaves a clear and distinct signature in the sound waves. The system was also able to estimate the maximum depth of the crack, though with slightly more variation. The average error for depth was about 2.10 millimeters. While this is a useful level of accuracy for a first estimate, the results showed that the exact deepest point of the crack was harder to pinpoint than the overall angle.

The study highlights a crucial distinction in how these results should be understood. The entire process was conducted using simulated data. The sound signals were not recorded from real steel blocks in a laboratory; they were generated by a physics-based software model called CIVA. This model is highly accurate for theoretical studies, but it does not account for the messy realities of the real world, such as imperfect connections between the probe and the metal, surface roughness, or electronic noise. Therefore, the success of the method proves that the direct geometric approach is feasible in theory and within the controlled environment of the simulation. It does not yet prove that the method will work perfectly on a real-world inspection site. The researchers explicitly state that the next step must be to test the system with actual physical measurements to see how it handles the unpredictability of real materials.

The findings offer a new way to think about how machines can assist human inspectors. Currently, the standard practice involves creating a visual map of the sound data and then having an expert interpret that map to find the flaw's dimensions. The new approach suggests that for a defined type of inspection and a specific family of flaws, the machine could skip the visual map entirely and go straight to the numbers that matter: the depth and the angle. This does not replace the need for human oversight or the established methods of image reconstruction. Instead, it offers a complementary tool that could provide a rapid, quantitative estimate of a flaw's geometry. If the computer's direct estimate agrees with the expert's visual interpretation, it adds confidence to the assessment. If they disagree, it flags the case for a closer look.

Ultimately, this work demonstrates that the relationship between a sound signal and a crack's shape can be learned directly by a computer. The study shows that under controlled conditions, the complex physics of sound waves contain enough information to reconstruct the geometry of a flaw without intermediate image processing. The results are a proof of concept, showing that the path from sound to shape is navigable. While the method is currently limited to the specific conditions of the simulation, it opens a door for future research where such direct geometric inversion could become a standard part of the toolkit for ensuring the safety of critical infrastructure. The path forward involves taking this digital success and testing it against the physical world, bridging the gap between a perfect simulation and the imperfect reality of industrial inspection.

Technical Summary: Direct Geometric Reconstruction of Flaws from Simulated Ultrasonic Inspection Data

Problem Statement
Quantitative ultrasonic flaw characterization traditionally relies on an indirect inference process: operators or algorithms first interpret spatially indexed A-scans or reconstructed intensity images (e.g., via Synthetic Aperture Focusing Technique, SAFT) to estimate geometric properties such as depth, orientation, and extent. This study addresses the hypothesis that, for a fixed conventional single-element pulse-echo inspection configuration and a constrained family of flaws, the mapping from ultrasonic response to geometric parameters can be formulated as a direct geometric inverse problem. The goal is to bypass the intermediate step of image reconstruction and interpretation, instead estimating an ordered flaw contour directly from the raw inspection data.

Methodology
The study employs a data-driven approach using physics-based simulation and supervised deep learning:

Data Generation: A dataset of 1,096 paired cases was generated using the CIVA ultrasonic simulation module. The simulations modeled a fixed inspection setup: a 2 MHz single-element rectangular probe on a 45° wedge inspecting a steel component. The flaws were defined as faceted, inner-surface-breaking profiles.
Input/Output Representation:
- Input: A standardized 256 × 256 single-channel B-scan representation of the simulated ultrasonic response.
- Target: An ordered vector of 11 (x, z) contour points defining the flaw profile. This representation allows for the direct derivation of engineering quantities without intermediate image segmentation.
Model Architecture: A compact convolutional regressor (implemented in TensorFlow/Keras) was designed to map the B-scan input to the 22-dimensional coordinate output. The architecture consists of two convolutional blocks (32 and 16 filters, respectively) followed by a fully connected layer and a linear output layer. The model is intentionally lightweight, as it estimates a low-dimensional geometric state rather than performing dense image synthesis.
Training and Optimization: The model was trained using Mean Squared Error (MSE) on the normalized coordinates. Crucially, model selection and hyperparameter tuning were not based solely on coordinate loss but on the Mean Absolute Error (MAE) of two derived physical quantities: maximum flaw depth and tilt angle. A grid search identified a unique non-dominated architecture that balanced performance on both metrics.
Validation: The dataset was split into 70% training and 30% held-out test sets. Performance was evaluated on the fixed test set using the derived depth and tilt metrics.

Key Results
The study demonstrates that direct geometric inversion is feasible within the constraints of the simulated domain:

Tilt Estimation: The model achieved a high coefficient of determination ( $R^2 \approx 0.886$ ) and a Mean Absolute Error (MAE) of approximately 6.30° for the flaw tilt angle.
Depth Estimation: The reconstruction of maximum flaw depth yielded an $R^2 \approx 0.716$ with an MAE of approximately 2.10 mm.
Error Analysis: The results indicate that global profile orientation (tilt) is recovered more consistently than local depth extrema. This is attributed to tilt being informed by the coordinated positions of multiple contour points, whereas depth depends on a single extremal location, making it more sensitive to localized coordinate errors.
Qualitative Performance: Visual inspection of test cases confirms that the reconstructed contours generally preserve the position and overall direction of the reference flaws, with deviations appearing primarily as local vertical offsets rather than geometric misalignment.

Significance and Claims
The paper positions this work as a proof of concept for task-oriented geometric inversion in conventional ultrasonic testing. Its primary contributions and claims are:

Methodological Shift: It demonstrates that for a defined inspection configuration and a constrained flaw family, the reconstruction target can be aligned directly with physical geometry rather than an intermediate intensity image.
Complementarity: The author explicitly states that this approach is complementary to, not a replacement for, conventional methods like SAFT. In a practical workflow, direct inversion could provide rapid quantitative estimates that serve as additional evidence alongside traditional operator interpretation.
Simulation Limitations: The study is modest regarding its scope. It acknowledges that the results are bounded by the assumptions of the CIVA forward model (e.g., idealized coupling, material homogeneity) and do not yet represent field performance. The reported errors characterize interpolation within a specific virtual inspection procedure.
Future Requirements: The paper defines the principal requirements for subsequent validation, emphasizing the need for experimental data to address real-world variabilities (surface condition, noise, microstructure) and the necessity of comparing the learned model against conventional SAFT-based sizing workflows in terms of accuracy, robustness, and uncertainty communication.

In summary, the study validates the feasibility of using a compact convolutional regressor to perform direct geometric reconstruction of faceted flaws from simulated ultrasonic data, establishing a foundation for future experimental verification and uncertainty quantification.

More like this