A University of Michigan team has received a $750,000 U.S. Department of Energy Genesis Mission award to develop AI agents that automate the design of custom computer chips for scientific computing. The project, called ArchEvolve, pairs U-M researchers with Los Alamos National Laboratory (LANL) and Intel to cut the years-long engineering cycle required to build application-specific accelerators.
Reetuparna Das, professor of computer science and engineering at U-M, leads the project with co-principal investigators Scott Mahlke and Seah Kim. The Phase 1 award is split between LANL and U-M.
Custom accelerators are chips built for specific workloads rather than general-purpose processing. They can deliver large gains in speed and efficiency, but only if the hardware matches the software it runs. That matching process, called hardware-software co-design, currently requires teams of human experts working through slow, iterative cycles that can take years.
Why chip design is slow
Scientific computing workloads are especially hard to accelerate. Many are irregular and memory-intensive, making them difficult to run efficiently on existing chips or GPUs. Engineers must weigh trade-offs across the entire computing stack, from the algorithm to the transistor layout, and small changes in one layer can force rework in others.
"The goal is to make accelerator chips' architecture design much easier and more automated using AI agents," said Das. "We want to reduce the engineering effort required to build specialized computing systems, especially for high-performance computing applications."
How ArchEvolve works
The team will build a multi-agent AI framework where different agents handle different parts of the design loop. One agent generates candidate hardware architectures. Another compiles and maps software to that hardware. A third evaluates designs for performance, efficiency, and feasibility.
The system is designed to incorporate expert knowledge into the architecture design process, supporting the development and refinement of specialized architectures rather than replacing human judgment entirely. The work builds on broader efforts to apply AI for Science & Research workflows to DOE priorities.
"Architecture design often requires repeated iteration across the computing stack. ArchEvolve aims to reduce that manual effort and help researchers develop promising designs more efficiently," said Kim.
The project also fits into the growing use of AI Agents & Automation for engineering tasks that have resisted automation because they require deep domain expertise.
From scientific code to custom hardware
If the approach works, a researcher could take new scientific code and rapidly explore custom accelerator designs that make it run faster. That capability could support DOE-relevant workloads in nuclear science, physics simulations, energy systems, and climate science.
"If we can shorten the path from scientific application to efficient hardware and software, we can help researchers rethink how specialized systems are designed and respond much more quickly to scientific challenges," said Mahlke.
The DOE Genesis Mission is a national initiative that pairs AI with supercomputing, quantum systems, and advanced scientific instruments. Phase 1 awards support projects exploring how AI-enabled workflows can improve experimentation, prediction, design, and discovery across science and energy research.
Why this matters for scientists and researchers
For researchers who depend on high-performance computing, the practical stakes are simple: today, getting custom hardware for a specific scientific workload can take years and requires rare expertise spanning hardware design, compilers, and application development. ArchEvolve targets that bottleneck directly. If AI agents can compress the design cycle from years to months, scientists could request accelerators tailored to their actual code instead of adapting their research to whatever general-purpose hardware is available. That would change how labs plan compute-heavy experiments and could make specialized systems practical for a wider range of scientific problems.
Your membership also unlocks:
