AI Method Boosts Metal 3D Printing Defect Detection Accuracy by 85 Percent

Addressing key challenges in additive manufacturing

Additive manufacturing allows the creation of complex components that are often impossible to fabricate using traditional methods. It also minimizes material waste, commonly measured as the buy-to-fly ratio, and reduces environmental impact. However, AM processes, particularly those involving metals, remain prone to defects due to the complex physical interactions that occur during material deposition. These flaws can lead to scrap, rework, and financial losses, making certification and inspection essential for structural and safety-critical parts.

Conventional non-destructive testing (NDT) methods, though reliable, are time-intensive and typically applied to random samples rather than full production runs. To improve this process, Dr. Giulio Mattera of the University of Naples “Federico II” (Italy), together with Professor Zengxi Pan of the University of Wollongong (Australia) and their collaborators, developed a new AI-based framework for real-time process monitoring and defect prevention.

AI-driven data analysis for welding-based additive manufacturing

“Data exhibit complex structures and relationships, especially in welding-based technologies,” explains Dr. Mattera. “Therefore, a more sophisticated analysis of the data, enabled by advanced data analytics and Artificial Intelligence, can outperform traditional methods based on simple statistical descriptors such as the mean and variance of the process.”

The proposed method processes high-frequency welding current and voltage signals, sampled at a rate of over 5,000 times per second, to capture patterns from both the time and frequency domains. This dual-domain analysis allows the model to detect not just numerical anomalies but also subtle structural deviations in the signals’ repetitive nature, which are closely linked to process stability and defect formation.

“A stable Wire Arc Direct Energy Deposition process produces repetitive waveforms in both current and voltage that correspond to the melting and deposition of filler wire,” notes Dr. Mattera. “By jointly analyzing information from both domains, we can better evaluate process stability and identify anomalous conditions that could lead to quality degradation.”

Minimal training data and adaptive learning

A key innovation of the methodology lies in its ability to operate with minimal prior data. “Unlike conventional AI models that require extensive datasets with both good and defective samples, our approach only needs data from non-defective conditions,” Dr. Mattera explains. “This drastically reduces data collection time and cost. The model learns to recognize normal behavior and automatically flags deviations, detecting defects it has never explicitly encountered.”

Experimental validation and comparative performance

In a series of controlled experiments, the team varied process parameters to gather datasets from high-quality deposition conditions. An Isolation Forest algorithm was then trained to establish a baseline of normal operational behavior. During testing, the model successfully identified deviations associated with process instability, including irregularities suggesting the need for maintenance interventions, such as cleaning the welding torch nozzle to prevent porosity defects in the deposited metal.

Compared with traditional Statistical Process Monitoring (SPM), the AI-driven method showed markedly higher accuracy. The SPM approach detected only 43 anomalies, missing or misclassifying 105. By contrast, the AI model correctly identified 116 anomalies while missing just 32, resulting in an improvement in anomaly detection performance from 57% to 85.3%.

“Our methodology can detect anomalous conditions that conventional approaches overlook,” said Dr. Mattera. “Current in-process monitoring methods are good at flagging extreme conditions but not at preventive maintenance, where early detection is critical to avoid defects before they form.”

Industrial implications and future development

These results underscore the potential of AI to transform process monitoring in advanced manufacturing. The team envisions applications extending beyond additive manufacturing to other welding-based industrial systems. “The results are very encouraging,” said Dr. Mattera. “We’re working to provide industry with practical tools that can enhance both productivity and quality.”

He added that the next phase of research will focus on integrating explainability and quality index estimation into the model to improve human-machine interaction. “We want to make these systems intuitive and collaborative, so operators can understand why an anomaly is flagged and take preventive action.”

Path toward industrial adoption

While this research represents a significant advance in AI-supported manufacturing qualification, regulatory and standardization frameworks remain crucial for the real-world implementation of this technology. “Our next goal is to refine and industrialize the methodology,” Dr. Mattera concluded. “By improving explainability and transparency, we can make AI a trusted partner in ensuring manufacturing quality and safety.”