Artificial intelligence (AI) has emerged to automate analog semiconductor design, which has heavily relied on human intuition and experience.
A research team led by Professor Kim Byeong-seop in the Department of Electrical Engineering at Pohang University of Science and Technology POSTECH said on the 29th that it developed AI technology that can overcome the limits of analog semiconductor design. The results were published in October last year in the international journal Institute of Electrical and Electronics Engineers (IEEE) Transactions on Circuits and Systems.
Semiconductors are core components that underpin modern industries, including smartphones, cars, and AI servers, but the design process still requires a lot of human labor. In particular, placement and routing (layout) work, which determines circuit performance and stability, must satisfy numerous design rules, so engineers often complete it by manually fitting the structures together.
Analog semiconductors, in particular, have more complex structures than digital ones and vary greatly in design methods by circuit, making automation itself difficult. In addition, because semiconductor design data are a corporations' core asset and disclosure to the outside is restricted, securing data needed for AI training has not been easy, which has been cited as a stumbling block to spreading the technology.
The solution the team focused on is a general-purpose AI—namely a "foundation model"—that first learns from large-scale data and then performs diverse tasks with only small amounts of additional training. The team devised a method to train AI to learn on its own the geometric pattern (layout) design required in the process of implementing analog circuits into actual semiconductor chips.
As a result, the team generated about 320,000 training samples based on six sets of real semiconductor design data. After pretraining, the AI learned the common structures and patterns that repeatedly appear in the design process and was able to perform the design needed for circuit consolidation and structure formation with relatively little additional data.
In validation experiments, 96.6% of the AI-generated designs passed both design-rule and circuit verification. The team said there is no need to build a separate AI model for each task as before, and that various analog design tasks can be expanded from a single foundation model, which can reduce the burden on design staff and help shorten development time.
Professor Kim Byeong-seop said, "This is a result that substantially broadened the possibilities of automating analog semiconductor design, which had been blocked by data scarcity."
References
IEEE Transactions on Circuits and Systems (2025), DOI: https://doi.org/10.1109/TCSI.2025.3615646