Designing high‑performance microwave absorbers traditionally demands expertise in electromagnetics, materials science, and simulation programming, and the optimization process can be very time‑consuming. AbsorbEvo introduces an agentic framework that translates natural‑language performance objectives directly into designs verified by full‑wave simulations.
The core of the framework is a candidate evolution strategy that fuses language reasoning, physics‑based prediction, and historical feedback. A large language model (LLM) proposes the direction and magnitude of parameter adjustments based on the task goal and computational history. The system then performs two types of sampling: directed increments guided by the LLM and global random sampling to maintain coverage of the design space.
All generated candidates are ranked by a low‑cost predictive model serving as a physics prior; only the top‑ranked designs proceed to full‑wave simulation. Results that pass physical validity checks are used to evaluate performance and feed back into the search, steering subsequent adjustments.
Experience gathered during training is distilled into textual skills, which are independently validated before being applied to new tasks, preventing negative transfer.
Under identical proposal budgets on 36 held‑out AbsorbBench tasks, AbsorbEvo achieved a 79.17% task success rate, far surpassing a generic agent (25.00%) and random search (12.50%). Its mean best coverage was 0.7816, compared with 0.6434 and 0.6448 respectively.
By integrating language reasoning with physics‑based feedback into design decisions, AbsorbEvo provides a methodological foundation for natural‑language‑driven autonomous inverse design of microwave absorbers.
Review: AbsorbEvo demonstrates the powerful synergy of LLMs and physics models for automated material and electromagnetic design, opening a promising avenue toward fully autonomous engineering workflows.