Porosity Prediction in Aluminum Wire Arc Additive Manufacturing Using Two-Stage Machine Learning Cascade and Computed Tomography

Imagine building a massive metal structure by melting a wire and laying it down, layer by layer, like a very fast, very precise 3D printer. This process, known as wire arc additive manufacturing, is changing how we make large metal parts for industries like aerospace. It is efficient and uses materials well, but it has a stubborn flaw: tiny air pockets, or pores, can get trapped inside the metal as it cools. These invisible holes weaken the structure, much like a crack in a dam, and can cause the part to fail later. For decades, engineers have had to wait until a part was finished to check for these flaws, often using heavy X-ray machines to look inside. If a defect was found, the entire expensive part might have to be scrapped. The goal for modern manufacturing is to catch these problems while the part is still being made, adjusting the process in real time to stop the holes from forming in the first place.

A team of researchers at Roketsan in Turkey has taken a significant step toward this goal by teaching a computer to predict when these pores will appear. They focused on a specific type of aluminum alloy, Al-2319, which is prized for its strength and is commonly used in rockets and aircraft. To understand the problem, they built several tall walls of this metal using a robotic arm and a specialized welding system that moves the wire and electric arc with extreme precision. As the robot worked, it recorded every tiny change in the welding process, such as how fast the wire was fed, the voltage of the electric arc, and the temperature, capturing this data every tenth of a second. Once the walls were finished, the researchers did not just look at the outside; they scanned the entire inside of each wall using industrial computed tomography. This is a powerful form of 3D X-ray imaging that creates a detailed map of every single pore inside the metal, showing exactly where they are, how big they are, and what shape they take.

The real challenge was connecting the dots between the split-second changes in the welding machine and the permanent holes that appeared inside the metal. The researchers had to align two very different sets of information: the continuous stream of welding data and the static 3D map of the pores. They built a system that matched the location of each pore to the exact moment in time when the welding parameters were recorded at that spot. This allowed them to see what the machine was doing just before a pore formed. By studying this data, they discovered that the formation of these defects was not random. Instead, specific patterns in the welding process often preceded the appearance of a hole. For instance, they found that when the wire feed speed increased slightly or when the voltage spiked, the likelihood of a pore forming went up. Interestingly, the temperature of the metal and the flow of shielding gas played a role, but the most telling signs were the subtle, rapid fluctuations in how the wire was fed and the electrical power was delivered.

To turn these observations into a practical tool, the team developed a two-stage computer system designed to act like a safety net. The first stage, which they call a "sentinel," is a simple but sensitive model that watches the welding data and flags anything that looks suspicious. Its only job is to catch as many potential problems as possible, even if it raises a few false alarms. The second stage, the "confirmer," takes only those flagged moments and examines them more closely using a more complex model. This second model is stricter; it tries to verify if the alarm is real, filtering out the false positives. The researchers tested this system on new walls that the computer had never seen before. While the system did not catch every single defect, it successfully identified a significant number of them and provided a stable stream of warnings that could be used in a real factory setting. They also found that smoothing out the warnings over a few seconds helped prevent the system from flashing alarms too quickly, which would be distracting for a human operator.

The study suggests that by watching the welding process closely and using this two-step computer check, manufacturers can move away from just inspecting parts after they are built. Instead, they can potentially adjust the machine while it is working to prevent defects from forming. The researchers noted that this approach relies on data that is already available from standard welding machines, meaning it does not require expensive new sensors like high-speed cameras or acoustic microphones. While the system is not perfect and still misses some defects, it offers a scalable and understandable way to improve quality control. The findings indicate that the key to stopping these invisible holes lies in managing the stability of the wire feed and the electrical arc, providing a clear path for making stronger, more reliable metal parts in the future.