AI Designs Computer Chips We Can't Understand — But They Work Really Well

chip-designinverse-designai-hardwareblack-box-ai

Abstraction: AI inverse design of wireless RF chips surpassing human performance

Key points:

Princeton and IIT Madras researchers used deep learning for "inverse design" — starting from desired properties to generate circuit geometry, published in Nature Communications
Convolutional neural networks (CNNs) trained to map circuit geometry to electromagnetic behavior for RF and sub-terahertz chips
AI-designed circuits are randomly shaped and counterintuitive, yet outperform classically designed ones with previously unachievable performance
Designed compact dual-frequency antennas and precise band-pass filters in minutes vs. days/weeks for human engineers
Lead researcher Kaushik Sengupta: "Humans cannot understand them, but they can work better" — raises black-box safety concerns
National Semiconductor Technology Center awarded ~$10M to continue this effort led by Princeton

Connections: Princeton University · Inverse Design · Convolutional Neural Networks · Explainability