Fermilab Genesis Mission: AI Silicon Design in Minutes

Engineers at Fermilab have achieved a breakthrough in semiconductor design with the unveiling of AXESS and TREASURE, two frameworks designed to produce radiation-hardened AI chips in minutes rather than months. Part of the Genesis Mission, these tools are set to revolutionize the way we design silicon for extreme environments.

Minutes-to-Silicon: The AXESS Framework

Traditional ASIC (Application-Specific Integrated Circuit) design is a laborious process that can take over a year. AXESS (Automated eXpansion of Silicon Systems) uses generative AI to automate the layout of transistors and interconnects. By inputting high-level functional requirements, AXESS can generate a GDSII file—the standard format for fab production—in under 60 minutes.

The framework is particularly adept at optimizing for low-power inference at the edge. For high-energy physics experiments at the Deep Underground Neutrino Experiment (DUNE), these chips must process terabytes of data locally to identify relevant events while surviving intense radiation bombardment.

TREASURE: Radiation-Hardened AI

Space is a hostile environment for standard electronics. Cosmic rays can cause Single Event Upsets (SEUs), leading to data corruption. TREASURE (Triple-Redundant Electronic Architecture for Space and Undersea Research Environments) introduces Triple Modular Redundancy (TMR) at the logic gate level automatically.

The TREASURE compiler analyzes the AI model's weights and activations, identifying critical paths that require additional protection. It then weaves in Error Correction Code (ECC) and redundant processing lanes without significantly bloating the chip's footprint. This allows the Genesis Mission to employ advanced computer vision models on its asteroid-scouting probes.

Impact on the Semiconductor Industry

Fermilab's work is being closely watched by the commercial space sector and the defense industry. The ability to rapidly iterate on rad-hard designs reduces the cost of space missions by orders of magnitude. Furthermore, the AXESS methodology could be applied to consumer electronics, enabling hyper-specialized chips for specific software tasks.

By open-sourcing parts of the AXESS codebase, Fermilab is encouraging a new generation of engineers to experiment with AI-assisted silicon design. The goal is to move toward a world where hardware is as agile as software, enabling the next frontier of discovery in both subatomic physics and deep space exploration.