AbsorbEvo: An Agentic Framework for Autonomous Inverse Design of Microwave Absorbers
Inverse design of microwave absorbers traditionally demands rare cross-disciplinary expertise and lengthy simulation cycles. AbsorbEvo tackles this with an autonomous agentic framework that translates natural-language performance goals into candidate designs, screens them with a low-cost physics prior, and verifies select candidates with full-wave simulations. The core idea is to marry large language model reasoning with electromagnetic physics feedback so that expensive simulations are spent only on promising candidates. If robustly validated, such a pipeline could lower barriers to designing broadband, angle-tolerant absorbers.
The paper presents a conceptually appealing multi-fidelity agentic pipeline, but the empirical case is weaker than advertised. The reported success rate of 79.17% on 24 held-out AbsorbBench-36 tasks is substantially above random and a handicapped generic-agent baseline, yet the no-feedback ablation—which retains the physics prior and fixed domain guidance but withholds new simulation feedback—already attains 62.50%. This suggests the LLM's closed-loop adaptation contributes only a modest increment over a well-informed, open-loop physics-screened search. More critically, the manuscript relies on undisclosed models ('gpt-6-luna', 'gpt-5.6-terra', 'gpt-6-astra') that are neither cited nor described, leaving readers unable to reproduce the results or assess whether the gains stem from the architecture or an unusually capable underlying model. The benchmark is also small, covering only two structural families, and the skill-transfer experiment is limited to eight held-out tasks.
The multi-fidelity optimization architecture is sound: candidate designs are generated via directed increments proposed by an LLM plus global sampling, then ranked with a low-cost effective-medium physics prior so that only the top candidate is submitted to expensive full-wave simulation. The physics prior shows respectable rank agreement with full-wave results, with Spearman correlations up to 0.978 for MAHSS and 0.935 for TPMS mean joint coverage. In addition, the deterministic physics-verification layer—checking power reconstruction, energy consistency, reciprocity, and solver convergence—provides a credible safeguard against accepting invalid simulation outputs.
A central flaw is the use of undisclosed model identifiers ('gpt-6-luna', 'gpt-5.6-terra', 'gpt-6-astra') without citations, API details, or prompt descriptions, which makes the work irreproducible and the baseline comparisons unverifiable. The benchmark scope is also very narrow: only 24 held-out tasks spanning two geometric families, with just five proposal rounds each, and the generic-agent baseline is deliberately deprived of the physics prior and domain guidance that AbsorbEvo enjoys, exaggerating the reported margin. Moreover, the ablation analysis reveals that most of AbsorbEvo's advantage can be attributed to the physics-prior screening and fixed domain guidance: the no-feedback variant still succeeds on 15 of 24 tasks (62.50%), meaning the LLM's iterative, history-guided adaptation adds only four tasks (16.67 percentage points). The skill-transfer claim is similarly tentative, resting on only eight held-out tasks and a validation-based selection procedure that risks overfitting the external skill to the validation set.
The evidence supports the narrower claim that a physics-informed screening step improves search efficiency within the tested task set, but it does not support broader claims of superiority over established inverse-design methods. The manuscript omits comparisons to classical optimizers such as genetic algorithms or Bayesian optimization, which are standard in electromagnetic topology and parameter optimization. The prior–exact correlation analysis pools historical simulation records rather than using an independent hold-out set, and the authors correctly acknowledge that theoretical predictions are therefore used for parameter screening, while design updates and target-attainment assessments remain based on full-wave results. Comparisons to related agentic frameworks in photonics are qualitative; no quantitative head-to-head results are provided against Lu et al.'s or Lupoiu et al.'s agents on shared tasks.
Reproducibility is severely compromised. The source code and benchmark data are not yet available—the authors state that the complete source code for this study will be made publicly available upon acceptance of the manuscript—and the LLMs identified as 'gpt-6-luna', 'gpt-5.6-terra', and 'gpt-6-astra' cannot be independently accessed or verified because no provider, version, API documentation, or hyperparameter snapshot is given. While the full-wave solver (CST Studio Suite 2025) and the physics-prior equations are documented in detail, the absence of the LLM orchestration layer, prompts, and the parametric geometry builders means independent reproduction of the agentic workflow is currently impossible.
Designing high-performance microwave absorbers requires specialized expertise in electromagnetic theory, materials science and simulation programming, and entails time-consuming optimization. Here, we present AbsorbEvo, an agentic framework for autonomous inverse design that translates natural-language performance objectives into designs verified by full-wave simulations. Its candidate evolution strategy integrates language reasoning, physics-based prediction and historical feedback. A large language model proposes the directions and magnitudes of parameter adjustments based on task objectives and computational history. The system combines directed increments with global sampling to generate candidates and uses a low-cost predictive model as a physics prior to rank them. Only high-ranking designs undergo full-wave simulation. Results passing physical validity checks are used to evaluate performance and guide subsequent search. Experience from training tasks is further distilled into textual skills, which are independently validated before use in new tasks. Under identical proposal budgets on held-out AbsorbBench-36 tasks, AbsorbEvo achieved a task success rate of 79.17%, versus 25.00% for a generic agent and 12.50% for random search. Its mean best coverage was 0.7816, compared with 0.6434 and 0.6448, respectively. By integrating language reasoning and physics-based feedback into design decisions, AbsorbEvo provides a methodological foundation for natural-language-driven autonomous inverse design of microwave absorbers.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.