AI automates analog semiconductor layout design, traditionally reliant on manual engineering/ hampered by lack of shared training data

https://techxplore.com/news/2026-02-ai-analog-layout.html

"unlike digital circuits, difficult to automate: high sensitivity, data scarcity (proprietary), design strategies vary significantly between different circuit types... self-supervised learning similar to LLM training... predicts missing layout parts looking at surrounding data patches... generated 320K training samples from 6 existing datasets... fine-tuned for contact/ via generation, dummy pattern insertion, N-well generation, metal routing. 96.6% accuracy, using only 12.5% data required by conventional... 1 model/ multiple tasks"