AI for Weld Quality Assessment: Detecting Porosity and Incomplete Fusion

Welding is one of the most common joining processes in manufacturing, and weld quality has direct implications for product safety and performance. A weld that looks perfect on the surface can contain internal defects that compromise its strength: porosity from trapped gas, incomplete fusion where the weld metal did not properly bond with the base material, or lack of penetration where the weld did not reach the full depth of the joint.

Traditional weld inspection relies on a combination of visual inspection, destructive testing of sample welds, and non-destructive testing (NDT) methods like X-ray, ultrasonic, or magnetic particle inspection. Each of these has limitations. AI is making all of them more effective.

In-Process Monitoring

The most valuable place to detect weld quality problems is during the welding process itself, when you can still do something about it. AI-based in-process monitoring analyzes the welding parameters in real time to predict whether the resulting weld will be good.

The data sources include welding current, voltage, wire feed speed, travel speed, shielding gas flow, and sometimes thermal imaging of the weld pool. These parameters have complex interactions. A slight drop in shielding gas flow combined with increased travel speed might produce porosity, while either change alone would be fine.

The AI learns these interactions from thousands of documented welds where the parameters were recorded alongside the quality outcome. It builds a model that predicts the probability of various defect types based on the parameter values it observes during welding.

When the model detects a combination of parameters that correlates with defects, it can alert the welder in real time to adjust technique, flag the weld for additional inspection, or in automated welding systems, adjust the parameters automatically.

Automated NDT Interpretation

Non-destructive testing generates data that requires expert interpretation. A radiographic image of a weld shows the internal structure, but reading it accurately requires training and experience. An ultrasonic scan produces signals that indicate reflectors inside the weld, but distinguishing a defect from a normal geometric feature takes skill.

AI trained on large datasets of NDT results with known outcomes can automate this interpretation. For radiographic (X-ray) inspection, computer vision models detect and classify defects including porosity, slag inclusions, cracks, incomplete fusion, and lack of penetration. The AI processes images faster than a human inspector and with more consistent sensitivity.

For ultrasonic testing, AI processes the A-scan, B-scan, or phased array data to identify reflectors, classify them as defects or benign features, and measure their size and location. This is particularly valuable for automated ultrasonic inspection of welds in pipe or structural steel, where the volume of data would overwhelm manual analysis.

Defect Classification and Severity Assessment

Detecting a defect is only the first step. The AI also needs to classify it and assess its severity. Different defect types have different implications for weld performance:

• Porosity reduces the effective cross-section of the weld. Scattered porosity at low levels might be acceptable under some codes, while clustered porosity usually is not.
• Incomplete fusion creates a planar defect that acts as a stress concentrator and can initiate cracking under load. It is almost always rejectable.
• Lack of penetration means the weld did not reach the root, leaving a gap that reduces joint strength and can act as a fatigue crack initiator.
• Slag inclusions are non-metallic material trapped in the weld. Small isolated inclusions may be acceptable, while linear or aligned inclusions are more serious.

The AI classifies defects according to the applicable welding code or standard, whether that is AWS D1.1 for structural steel, ASME Section IX for pressure vessels, or ISO 5817 for general weld quality. This automated code compliance assessment dramatically reduces the time and expertise needed for weld quality evaluation.

Process Feedback Loop

The real power of AI weld quality assessment comes from closing the loop between inspection results and welding parameters. When the AI detects a recurring defect pattern, it can trace the root cause back to specific process variables.

If porosity is increasing on a particular welding station, the AI might identify that the shielding gas flow has dropped slightly, or that the contact tip is worn and causing erratic arc behavior, or that the base material from a specific lot has higher sulfur content. This root cause identification is what turns quality inspection from a rejection mechanism into a process improvement tool.

For a broader look at AI quality systems in manufacturing, visit the FirmAdapt manufacturing analysis page.