Semiconductor chips—tiny, intricate discs used in electrical and photonic devices—are built through a highly complex, multilayered production process that is among the most sophisticated in manufacturing, according to Zheng Yao, principal research scientist in Lehigh’s Energy Research Center. Quality control during this process is essential.
Yao has partnered with Broadcom, which manufactures the chips in a Lehigh Valley facility, on a research project aimed at improving quality-control practices on the production line. Currently, engineers manually inspect data and microscopic images of semiconductors, a process that Yao calls “time-consuming and also very subjective.” If two or more quality-control engineers have different thoughts on one product, time can be lost resolving the issue, he explains. With tens of thousands of chips produced every day, this method can significantly delay operations.
Yao’s team is introducing machine learning to transform this process. Their project, funded in part by the Pennsylvania Infrastructure Technology Alliance (PITA), uses artificial intelligence to analyze data, classify images, and assess product quality without interrupting production. Early results have been promising, and Broadcom has expressed interest in scaling the tools.
Collaborating on this project are two assistant professors in the Department of Computer Science and Engineering. Yu Yang specializes in mobile sensing, spatio-temporal machine learning, and reinforcement learning, bridging human and cyber technologies. Lifang He focuses on machine learning, data mining, and biomedical informatics.
Yao and He are also working together on a separate AI-driven project tackling the challenge of integrating “legacy machines” used in manufacturing processes into the Internet of Things era. These stand-alone devices lack connectivity and often have complex menus, creating inefficiencies in monitoring and optimizing their performance. The team’s approach uses machine learning and image analysis to refine images taken during the manufacturing process, extract data, and enable real-time monitoring. The system can discern detailed information—such as color, symbols, and letters—from imperfect images, ensuring that legacy equipment can operate efficiently without immediate upgrades.
Graduate students play a pivotal role in both projects, which provide opportunities to apply their classroom-acquired knowledge to real-world challenges. “We are helping the students to develop a way of thinking and exploring independently in the future,” says Yao.