Machine learning and AI-supported software systems can be used to improve the speed and accuracy of quality control in the production of electronic circuit boards. A specially developed deep learning software requires only a few images to adapt the inspection system to the respective application.

The size and competitiveness of the mobile phone industry has driven innovation in many industries, from imaging to software to even metallurgy. However, the semiconductor market has benefited the most. The demand for higher performance in smaller packages has been unrelenting for few decades. Apple released its latest Iphones, some powered by its new A17 Bionic chips, built on TSMC’s new 3 nm manufacturing process. These chips are smaller, faster and more power-efficient than their 5 nm predecessors. Each chip is said to features 19 billion transistors, some so small that their elements are only 12 silicon atoms wide.

According to reports, Apple will switch to a new circuit board material: RCC foils (resin coated copper). This allows the circuit boards (PCB) to be designed even thinner. However, RCC foils are very delicate, especially in the laminating process, the material is vulnerable to heat and pressure. The more complicated the manufacturing process becomes, the more important optical inspection becomes to check the quality.

Quality Control: Bottleneck in Production

Quality control is the major bottleneck in the printed circuit board manufacturing chain, including reliability testing and reworking defective boards. Improving the efficacy of quality control can significantly increase the production yield and throughput, reducing manufacturing costs and waste.

Most PCB manufacturers use automatic optical inspection (AOI) to monitor defects in their boards. This delivers strong results when there are defects in soldering, connections, pads and traces on printed boards. AOI also proves to be very useful for the early detection of problems that happen during assembly, such as shorts, open circuits, thinning soldering or scratches on traces. In particular, scratches can be fatal to a board, changing its electrical properties and causing a total malfunction of the completed product.

AOI has the advantage of being included directly at the end of the PCB production line, before lamination and etching, detecting possible defects earlier than other methods. Imaging systems capture high-resolution images, with resolution down to a few microns, and then comparing them to images of a perfect model board (also known as the golden board) or with an image database of both acceptable and defective samples.

Other than performing tests on the PCB under assembly, the AOI method can monitor the manufacturing process itself. Pick-and-place machines can respond to detected defects in real-time, correcting assembly defects like component misplacement and misalignment.

Limits of Rule-based Image Processing

Still, with the demand for smaller, higher-performance parts, the resulting complexities and subtleties in material faults means that traditional manual inspection or rule-based imaging may simply not be up to the task. One semiconductor OEM needed to detect a variety of fine defects on PCB components, including breakage, abrasion, contamination, fragments, and air bubbles. However, using traditional rule-based image processing was not providing the accuracy they required. They were facing an increase in defective parts that went undetected in their existing process, driving up costs. They needed a new solution.

To overcome these obstacles, the OEM decided to explore machine learning to meet their accuracy requirements for detecting defects on PCBs and their components. They chose Teledyne’s Sapera AI inspection software suite which features the Astrocyte AI training tool. The Sapera AI software allowed them to expand on their rule-based algorithms with AI functions within their AOI machine. The Sapera AI software turned out to be a suitable solution for the OEM, allowing them to use much of their existing system while providing more accurate detection of the subtle defects that other methods would miss.

Using Sapera AI, the OEM was able to achieve 98% accuracy in continual classification with 12-14 ms speed for 200 images and 100% accuracy with 453 good and 11 bad images. Additionally, they were able to achieve 99.62% accuracy with 259 images and 20 ms speed for object detection when looking for multiple defects on a part image at the same time.

Improvement in Defect Detection

This is representative of the huge improvements that have been made in machine learning over the past few years. While an AI system typically needed to be trained from scratch, requiring hundreds of image samples. However, today’s deep learning software is often pre-trained, so users may only need tens of additional samples to adapt the system to their specific application.

The result for this OEM was a production line that could accurately detect subtle defects on PCBs without the need for labor intensive manual inspection. The AI functions provided a reliable and consistent alternative to traditional rule-based image processing which had previously been unreliable in detecting subtle defects.

Overall, the OEM experienced improvements in both accuracy and speed of defect detection on PCBs because of Teledyne’s Sapera AI software, allowing them to reduce fallout while providing higher quality products that met their specifications.

Outlook

Today, the industry is still recovering from the worldwide semiconductor shortage that began in 2021. While analysts predict that nearly 70 percent of growth in semiconductors over the next decade will be driven by just three industries: automotive, data storage, and wireless communications, these industries are still playing catch-up from missed product launches, delayed rollouts, and higher prices. The pressure is on.

Machine learning and AI-informed software systems can improve speed and accuracy in their biggest bottleneck: quality control. Instead of a problem, companies can turn quality control into a competitive advantage, improving speed and lowering costs.

Author
Bruno Ménard, Software Director at Teledyne Dalsa