An AI agent can now design microwave absorbers just by reading a plain-English brief of what you want them to do.
The system, called AbsorbEvo, pairs a large language model with physics-based prediction and feedback from past attempts. The LLM proposes how to tweak a design's parameters, a cheap predictive model screens the resulting candidates, and only the most promising ones get run through full-wave electromagnetic simulation. Designs that pass physical-validity checks feed back into the search, and lessons learned get distilled into reusable text skills for future tasks. On a held-out benchmark of 36 absorber-design tasks, AbsorbEvo hit a 79.17% success rate, versus 25.00% for a generic AI agent and 12.50% for random search, with notably better coverage of the target performance too.
That gap matters because absorber design has long been a slow, expertise-gated process, requiring fluency in electromagnetic theory, materials science, and simulation tooling most engineers don't have. Folding that expertise into an agent that takes natural-language specs could shrink a multi-week optimization cycle into something closer to a conversation, and the same pipeline looks adaptable to other inverse-design problems like antennas or metamaterials.
Still, this is one lab's benchmark built on its own tasks, not a product tested against real-world fabrication constraints, so treat the 79% success rate as a promising lab result rather than proof the hardware works outside a simulator.