Across sectors such as semiconductors, energy, automotive manufacturing, and steel production, the stability of equipment and structures depends on the flawless integrity of their internal components. Even microscopic cracks or defects—often hidden deep within materials—can have serious and sometimes catastrophic consequences. Because these flaws cannot be observed directly, engineers rely on non-destructive testing techniques to examine internal conditions without harming the structure. Although these methods have advanced considerably over the years, accurately detecting and characterising minor or complex internal defects remains difficult, as sensor signals are frequently distorted by geometric features, material irregularities, and the complex environments in which real systems operate.

These physical limitations have led researchers to ask a compelling question: could artificial intelligence reveal what human inspection tools cannot? This question has driven a significant innovation from a research team in South Korea led by Assistant Professor Sooyoung Lee of Chung-Ang University. Professor Lee, who heads the Industrial Artificial Intelligence Laboratory in the School of Mechanical Engineering, and his team have developed a diffusion-based AI system named DiffectNet. This model is designed to produce high-fidelity ultrasonic images that reveal hidden internal defects with far greater clarity than traditional signal interpretation techniques. Their findings were released online in September 2025 and subsequently published in the journal Mechanical Systems and Signal Processing.

DiffectNet represents more than an incremental improvement to existing methods. According to Professor Lee, it marks a fundamental shift in how AI can address engineering challenges. By learning to reason from complex, noisy sensor data, the system can reconstruct the probable shapes and locations of cracks or flaws within structures in real time. This generative ability allows the AI to overcome long-standing physical constraints associated with signal distortion, enabling a more precise and reliable view of internal conditions than was previously thought possible. In Professor Lee’s view, such capabilities have the potential to redefine safety standards across numerous industries.

The practical implications are far-reaching. In high-risk environments such as power plants, where even a tiny crack might lead to devastating accidents, AI-based real-time inspection could provide crucial early warnings. In semiconductor fabrication facilities, where production cannot easily be paused for detailed inspections, DiffectNet could allow virtual assessment of internal defects without interrupting equipment operation. This would support both improved quality control and sustained productivity. Similar benefits apply to the monitoring of critical infrastructure—bridges, tunnels, buildings, and transport networks—where concealed deterioration often goes unnoticed until it becomes severe. AI-driven reconstruction enables continuous, high-resolution evaluation, contributing to safer, more resilient urban environments.

This research illustrates the broader evolution of AI within engineering. No longer confined to data analysis, AI systems are becoming active participants in the discovery and interpretation of physical phenomena. Professor Lee emphasises that his team’s work is part of a larger movement toward AI-driven engineering technologies that can reshape the field’s very foundations. As such systems mature, they may transform how industries predict failures, maintain essential infrastructure, and ensure everyday safety.

Ultimately, DiffectNet offers a promising advance in safeguarding the reliability of the systems upon which modern life depends. Through its ability to reveal what was once hidden, this breakthrough points to a future where engineering decisions are guided by unprecedented insight and precision.

More information: Dongwon Lee et al, DiffectNet: diffusion-enabled conditional target generation of internal defects in ultrasonic non-destructive testing, Mechanical Systems and Signal Processing. DOI: 10.1016/j.ymssp.2025.113454

Journal information: Mechanical Systems and Signal Processing Provided by Chung Ang University