New AI Models: Early Detection of Microcracks and Material Defects in Manufacturing

Even the smallest microcracks can lead to massive production losses and serious safety risks in industrial manufacturing if left undetected. Early detection is critical in areas where material reliability is non-negotiable. Once again, AI proves to be a real game changer: when used effectively, it can identify hairline defects before they cause disruptions. Based on intensive research, our machine learning expert Fatahlla Moreh has developed two pioneering AI models, which he recently presented at the IEEE International Conference on Machine Learning and Applications (ICMLA) in Miami.

In this interview, Sascha Mahmood (Senior Specialist, Cloud Computing) and Fatahlla Moreh (Specialist Machine Learning Engineer) discuss how companies can leverage these models to redefine microcrack detection—and where the greatest practical value lies.

Sascha Mahmood: Fatahlla, what makes the two new AI models for damage detection so special?

Fatahlla Moreh: The new models MicroCracksMetaNet50E and MicroCracksAttNet50E are special because they use numerical wave data to detect and segment microcracks, instead of relying on image data. This method enables faster and more efficient analysis compared to traditional image-based methods. What is new is the use of deep learning techniques, specifically Meta's Segment Anything Model (SAM) for MicroCracksMetaNet50E, which achieves high accuracy (86.7%) and precision (90.6%) in crack detection. MicroCracksAttNet50E goes one step further and utilizes an attention mechanism that focuses on critical areas to identify even the smallest cracks more precisely. These innovations provide a cost-effective and resource-efficient solution for damage detection in areas such as material testing and structural health monitoring.

Sascha Mahmood: What methods make the models so effective and how are they used in damage detection?

Fatahlla Moreh: Microcracks in materials can cause serious damage if they remain undetected. Our deep learning models use numerical wave data to precisely detect these cracks – without the need for classical image processing. Instead, they directly analyze the numerical patterns that arise from material defects.

Our models combine convolutional neural networks (CNNs) with transformers to capture both local and global features of the wave data. This enables them to reliably segment even the smallest cracks.

• MicroCracksMetaNet50E is based on Meta's Segment Anything Model (SAM). It uses CNNs for feature extraction, while the mask decoder works with transformer mechanisms to generate precise segmentations.
• MicroCracksAttNet50E additionally relies on an attention mechanism that focuses on critical areas and thus also efficiently detects very fine cracks.
  
  

This technology is ideal for:

• Material inspection: Automatic detection of the smallest defects during production.
• Structural health monitoring: Continuous monitoring of structures to identify damage at an early stage.

Thanks to this innovative combination of CNNs and transformers, our models work faster, more accurately and more efficiently – a decisive advantage for safety-critical industries.

Sascha Mahmood: What challenges have previous models been unable to overcome and how do the new models approach these problems?

Fatahlla Moreh: The existing YOLO models face several challenges when detecting micro-cracks. On the one hand, they divide images into grid cells, which causes problems when a crack is very thin or complex. Such cracks can be distributed across several cells, making it difficult to determine their exact shape, thickness and structure. This leads to inaccurate results, especially for fine or irregular cracks. Furthermore, YOLO models are primarily optimized for the detection of larger objects and have difficulties reliably identifying small microcracks.

MicroCracksMetaNet50E and MicroCracksAttNet50E address these problems directly. They work with numerical wave data that provides more precise information about the structure of cracks without being constrained by the limitations of a grid. In particular, MicroCracksAttNet50E uses an attention mechanism that focuses on the relevant areas and can thus accurately detect even very small and complex cracks. This approach enables greater accuracy in damage detection and is also more resource-efficient, since less computing power is required than with image-based methods.

Sascha Mahmood: Where is the greatest potential for application in practice?

Fatahlla Moreh: The new models offer great potential in manufacturing and in structural health monitoring (SHM). In manufacturing, they enable the early detection of microcracks in materials, which helps to avoid production errors and reduce material costs. In SHM, the models can continuously monitor structures such as bridges, buildings or aircraft and detect microcracks before they cause serious damage. This enables more efficient maintenance, saves costs and extends the service life of structures and machines. These models thus offer significant added value through precise, resource-saving and cost-effective damage detection.

Sascha Mahmood: So the models score highly in terms of accuracy and precision. In which area of application is this particularly important?

Fatahlla Moreh: In practice, the high accuracy and precision of the models means that microcracks can be detected with a high degree of reliability, which is particularly important in safety-critical areas such as bridge inspection. For example, if a microcrack in a bridge is overlooked, it could lead to serious structural damage over time. The models detect such cracks at an early stage, enabling targeted maintenance measures to be taken directly. This not only prevents increasingly costly repairs, but also reduces the risk of accidents. The high level of precision ensures that false reports are minimized and only relevant damage is reported, which further increases efficiency and cost control.

Sascha Mahmood: How can companies use the new models to meet their specific challenges?

Fatahlla Moreh: Companies can adapt the MicroCracksMetaNet50E and MicroCracksAttNet50E models to their needs relatively easily because they are based on numerical wave data. The key to success is to provide real, diverse data that covers as many edge cases as possible. This data is important to make the model more robust and to deliver accurate results even in complex situations. Companies can further customize and improve the model by fine-tuning or retraining it with their own data. After training, the model can be easily integrated into existing systems and used in areas such as manufacturing or structural health monitoring. This way, companies benefit from precise damage detection and flexible application of the models.

Do you want to improve your materials analysis?