AI Creates Radio Chip Designs That Humans Couldn't Even Imagine
Princeton University (Sengupta Lab)
Researchers at Princeton have taught AI to design radio-frequency integrated circuits (RFICs) using reinforcement learning, reverse engineering, and generative models. The new approach allows chips with record performance to be created in hours instead of months, with some designs resembling abstract art yet outperforming human-made counterparts.
A group of researchers at Princeton University led by Professor Kaushik Sengupta has developed machine learning methods for designing radio-frequency integrated circuits (RFICs), which are key components in 5G devices, autonomous vehicles, and satellite communications. Traditionally, RFICs are considered a 'dark art': their design requires years of experience and takes years to complete, with development costs reaching tens and hundreds of millions of dollars. The researchers applied reinforcement learning and inverse design to automatically synthesize RFIC topologies from scratch, without relying on pre-existing templates. Additionally, they used diffusion models capable of generating unusual yet efficient electromagnetic structures that outperform the best human-designed counterparts. Design time has been reduced by orders of magnitude, from months to hours. However, further progress requires open datasets and ecosystems so that artificial intelligence can learn universal laws of electromagnetism and circuit behavior.
Source: IEEE Spectrum AI —
original