Polymer manufacturers gain AI-powered tool to spot defects early

We rely on flexible plastic parts for everything from consumer goods and medical devices to automobiles. But for polymer manufacturers, ensuring consistent product quality can be challenging because the final performance of a part depends not only on the material and process parameters, but also on the environment in which it is made.

Temperature changes, humidity, vibration, airflow, dust and other contaminants can all affect how material flows and solidifies — and how defects form during fabrication. In polymer 3D printing, moisture is an especially persistent challenge, causing defects that often go undetected until after a part has been printed or, in some cases, breaks during use.

Now, a team of UC Berkeley-led researchers may have found a solution to this longstanding issue.

By combining artificial intelligence with inexpensive optical sensors, researchers have created a set of “smart eyes” for 3D printers that can spot moisture damage in real time. This simple, cost-effective monitoring system allows manufacturers to instantly identify defects that can affect part performance and intervene before production is complete.

Their findings were reported today in the journal Advanced Science.

“This technology is a major step forward for sustainable and reliable manufacturing,” said Grace Gu, associate professor of mechanical engineering and the study’s principal investigator. “If a 3D printer detects a critical flaw during printing, it can automatically alert the user, pause the process or eventually adjust the printing conditions. This prevents the waste of expensive materials and the energy costs associated with finishing a part that may later fail.”

Polymer manufacturers commonly use a flexible, rubber-like plastic called thermoplastic polyurethane (TPU) in 3D printing. While known for its high elasticity, shock absorption and durability, TPU also easily absorbs moisture from the surrounding air. When this “wet” plastic hits the hot printer nozzle, the trapped water instantly vaporizes, creating tiny bubbles, rough surfaces and weak spots in the final product.

Based on their tests, Gu’s team found that even a small amount of moisture or humidity can substantially reduce the physical strength of TPU plastic, causing nearly a 20% decrease in mechanical properties.

Until now, said Gu, manufacturers have been trying to ensure quality by using a pre-drying oven or through destructive testing to examine an object’s properties after production. Her team’s vision-based, real-time monitoring system will enable manufacturers to proactively evaluate the structural integrity of the object as it is being built.

To build their solution, the researchers equipped a standard 3D printer with a simple, low-cost camera and an LED light right next to the print nozzle. This allowed them to continuously watch the material being deposited in real time.

As explained by Jiyoung Jung, co-lead author and a postdoctoral researcher in the Gu Research Group, the “brain” behind this camera is an AI framework called a diffusion model.

“Instead of trying to teach the AI every possible way a print could fail, we only showed it what a ‘perfect,’ dry print looks like,” said Jung. “Using these dry print images, the model is trained to reconstruct the original image after noise has been added to it. After training, when the model sees a new image during printing, it tries to reconstruct it as if it were a normal dry print.”

He noted that if the print surface is normal, the reconstruction is accurate. However, if the surface has moisture-induced defects, such as roughness or bubbles, the model struggles to reconstruct those abnormal features.

The researchers measure that “struggle” and use it as a live warning signal. The system can then detect whether moisture-related damage is present and estimate how severe it is, processing each image in less than two-tenths of a second.

“This drastically enhances reliability because it evaluates the severity of the damage in real time, without damaging the print,” said Jung. “In this way, it acts as an instant quality-control checkpoint.”

This vision-based, real-time monitoring system for additive manufacturing detects subtle moisture-induced degradation via a diffusion model-based framework. (Image courtesy of the Gu Research Group)

“But the most exciting discovery was how incredibly sensitive our real-time monitoring system is,” said Yuna Yoo, co-lead author and a Ph.D. student in the Department of Mechanical Engineering. “We ran a test where we fused dry plastic and wet plastic together into a single continuous thread. Our AI successfully caught the moment the print shifted from ‘good’ to ‘bad’ — and vice versa — right in the middle of the print job.”

According to Yoo, the researchers were also surprised to learn that their model is highly adaptable. For example, they had only trained the system to recognize yellow-colored TPU plastic, but it was “smart enough” to detect moisture damage in nylon, a completely different, translucent material.

While 3D printing is currently used for functional applications such as soft robotics, customized medical devices, wearable systems and industrial polymer components, the researchers point out that its broader adoption will depend largely on manufacturers’ ability to make printed parts more reliable and consistent.

“By detecting environmental effects such as moisture and connecting those effects with final part performance, our system can help manufacturers identify defects early, reduce failed builds and eventually adjust the printing process before quality is compromised,” said Gu. “This could help expand the use of 3D printing in functional applications where mechanical performance, consistency and reliability are critical.”

Postdoctoral researcher Jiyoung Jung (right) and graduate student Yuna Yoo (left) from the Gu Research Group. (Image courtesy of the Gu Research Group)

The team sees this work as a starting point for studying how real-world conditions affect not only 3D printing, but also other advanced manufacturing processes where small changes can influence final product quality. In future studies, they plan to examine conditions beyond moisture, such as vibration, temperature and gravitational effects. These efforts could help manufacturers better understand and control the links among the environment, processing, material structure and final performance.

In addition to Gu, Jung and Yoo, co-authors include Dharneedar Ravichandran and Dahyun Daniel Lim, both from the Department of Mechanical Engineering. Lim is also affiliated with Korea University, Seoul, Republic of Korea.

This research was supported by Amazon and the Air Force Office of Scientific Research.