Page 158 - Read Online

Page 2 of 19 Wang et al. J. Mater. Inf. 2026, 6, 13

INTRODUCTION
Aerospace alloys including Titanium alloys and superalloys are often designed with complex geometries for
optimum performance and working under extreme conditions. A typical example is aero-engines blades,
which are designed with free-form surfaces for optimum aerodynamic performance [Figure 1]. As they are
usually working under very high pressure and rotational speed, minor surface defects may cause unexpected
fracture and threaten flight safety . Therefore, it is important to effectively detect defects such as scratches
[1-4]
and dents and evaluate their potential impact on the part lifespan.

It is still common to manually inspect the parts and identify the defects with naked eyes in the industry,
which is not reliable under fatigue . Meanwhile, the inspectors can only provide a qualitative judgment on
[5,6]
the defect existence, missing the quantitative measurement to evaluate further treatment. With the rapid
development on artificial intelligence in the material area [7,8] , deep learning approaches have shown
promising performance on the automated visual inspection (AVI) of aerospace alloy surface [9-12] . The
convolutional neural networks (CNNs) are one of the most popular structures [13,14] . Semantic guidance has
also been introduced to improve the detection performance [5,15] . The graph neural network is found useful for
the identification of irregular defects . Transformer-based networks have also been applied for this problem
[16]
and show promising performance . However, these studies defined defect detection as an object detection
[17]
problem, where the contour of defects is ignored.

As the aerospace alloys are expensive, and their working condition is severe, it is crucial to accurately
calculate the geometrical parameters of defects to evaluate their impact and inform treatment decisions [18-20] .
A transformer-based network for defect segmentation on aero-engine blade is proposed, showing the
potential to obtain the accurate shape of the defects for further evaluation . However, segmentation tasks
[6]
usually have a high requirement for training datasets. For the complex situation such as aero-engine blades
[Figure 2], it is very difficult to obtain comprehensive data that covers the possible defect presentation
distributions. In manufacturing stage, the defects are rare, making it more challenging to train a robust
network. How to tackle this situation is still an open question for aerospace alloy inspection.

The performance of deep learning-based models highly depends on the training datasets. However, data
scarcity remains a main challenge for complex parts . First, the curvature may change frequently all over
[10]
the surface . As a result, the same scratch may appear completely different in different areas and under
[21]
varying lighting and camera settings [Figure 2]. Combining the variety of scratch shapes, the possible image
distribution space is very huge. It is difficult to capture comprehensive datasets to cover this distribution.
Meanwhile, the defect is still a rare situation in practice, especially in the manufacturing process. For the
segmentation problems, labeling is also very time-consuming and difficult to maintain high quality .
[11]
Synthetic image generation is one of the main strategies to improve data quality and quantity for AVI.
Data-driven methods have dominated this area in recent years. Generative adversarial networks (GANs) are
among the most widely used methods [22,23] . More recently, diffusion model-based approaches have been
emerging rapidly . However, the development of these models requires high-quality data, and the generated
[24]
image distribution still follows the existing real datasets. Therefore, the model’s performance can easily
saturate as the volume of generated data increases . For defect detection of aerospace alloys, as their
[25]
distribution is very broad, adding possible image presentations outside the distribution of the existing
datasets is critical to improve the robustness of the segmentation.