A Self-Evolving Agentic Framework for Metasurface Inverse Design

Yi Huang; Bowen Zheng; Yunxi Dong; Hong Tang; Huan Zhao; S. M. Rakibul Hasan Shawon; Hualiang Zhang

cs.AIApr 1, 2026

A Self-Evolving Agentic Framework for Metasurface Inverse Design

Authors: Yi Huang, Bowen Zheng, Yunxi Dong, Hong Tang, Huan Zhao, S. M. Rakibul Hasan Shawon, Hualiang Zhang

Organizations: Department of Electrical and Computer Engineering, University of Massachusetts Lowell

Abstract

Metasurface inverse design can realize complex optical functionality, but turning a target optical response into executable optimization code still requires substantial expertise in computational electromagnetics and solver-specific software engineering. We present a self-evolving agentic framework that lowers this barrier by coupling a coding agent, explicit human-readable skill files, and a deterministic physics-based evaluator. Rather than updating model weights, it revises the skill files from solver-grounded feedback, while the base model and differentiable solver, which provides the physics simulation and gradients, stay fixed. On a multi-type benchmark, skill evolution raises same-type task success from 38% to 74%, the fraction of physical criteria met from 0.51 to 0.87, and reduces average attempts from 4.10 to 2.30. On two new-type families, success holds near ceiling on one (0.92 to 0.90) and rises from 0.20 to 0.90 on the other. Skill evolution offers a practical path toward autonomous and accessible inverse-design workflows.

Figures & tables

Figure 1: Overview of the self-evolving agentic framework for metasurface inverse design. The framework is organized as two decoupled layers: an upper code-generation layer, where a coding agent turns a natural-language request into a candidate optimization program guided by an evolving skill file ( SKILL.md ), and a lower physics layer, where a fixed differentiable solver performs the electromagnetic forward simulation and provides gradients, and an evaluator scores the result against the task criteria. The agent programs the fixed solver rather than replacing it, and the loop runs autonomously without a human in the loop. An inner loop retries the program when an evaluation fails, while an outer loop lets a meta-agent revise the skill file from per-task feedback, keeping the best version on validation; a passing design is released as the realized metasurface. The demonstrated solver is a differentiable RCWA implementation, TorchRDIT [ 8 ] .

Dataset	Condition	SG	SE	CPF	BM	Attempts
Same-type	Starter-skill baseline	38.0%	86.0%	0.510	-2.380	4.10
Same-type	Post-train	74.0%	100.0%	0.870	-2.281	2.30
Same-type	Delta	+36.0pp	+14.0pp	+0.360	+0.099	-1.80
New-type-A	Starter-skill baseline	92.0%	100.0%	0.960	-4.626	1.60
New-type-A	Post-train	90.0%	100.0%	0.950	-2.092	1.68
New-type-A	Delta	-2.0pp	+0.0pp	-0.010	+2.534	+0.08

Table 1: Test set performance: baseline (starter skill file) vs. post-train (evolved skills).

Figure 2: Outcomes of skill evolution across the three settings (Same-type, New-type-A, New-type-B), on held-out test tasks. (a) Self-evolution gain: each line traces SG from the starter skill to the evolved skill, with signed delta labels in percentage points. (b) Rescued vs. regressed: per setting, the number of tasks moving from failing to passing (teal, rescued) or passing to failing (red, regressed); the faded grey pool is the baseline-failing denominator. (c) Cost per task: box plots of total LLM spend per task (USD) for the starter and evolved skill in each setting.

Figure 3: Mechanism of skill evolution, one panel per setting (Same-type, New-type-A, New-type-B). Within each panel, the blue line is the validation success rate ( SG on the held-out validation family, right axis) across skill-evolution iterations; the skill evolution band shows the elitist search, where each node is a candidate skill file, an arrow marks a revision of the current best skill, and the star marks the iteration whose skill stays best on validation and becomes the final evolved skill; the stacked bars are the training-time error composition per iteration (left axis), with bar totals annotated. Because the loop keeps the best-on-validation skill and discards regressions, a later high-error candidate is a rejected exploration step, not a regression in the final skill. Infrastructure failures such as connection errors and timeouts are excluded. For New-type-B the seed marks the zero-shot baseline on the test families (its iteration 0 was scored there), whereas iterations 1 to 4 are on the validation family.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Family	Problem type	Design space	Criterion form
$G_{1}$	Single-condition reflective design	Low-dimensional periodic reflective geometry	$R(\lambda_{0};x)\geq\tau_{R}$
$G_{2}$	Dual-angle reflective design	Low-dimensional periodic reflective geometry	$\begin{aligned} R(\lambda_{0},\theta_{1};x)&\geq\tau_{R},\\ R(\lambda_{0},\theta_{2};x)&\geq\tau_{R}\end{aligned}$
$G_{3}$	Polarization-selective reflective design	Low-dimensional periodic reflective geometry	$\begin{aligned} R_{\mathrm{TE}}(\lambda_{0};x)&\geq\tau_{\mathrm{hi}},\\ R_{\mathrm{TM}}(\lambda_{0};x)&\leq\tau_{\mathrm{lo}}\end{aligned}$
$G_{4}$	Dual-wavelength reflective design	Low-dimensional periodic reflective geometry	$\begin{aligned} R(\lambda_{1};x)&\geq\tau_{R},\\ R(\lambda_{2};x)&\geq\tau_{R}\end{aligned}$
$G_{5}$	Plasmonic reflective design	Periodic reflective geometry in a plasmonic regime	$R(\lambda_{0};x)\geq\tau_{R}$
$G_{6}$	Transmissive phase-target metasurface design	Unit-cell geometry with family-dependent parameterization	$\begin{aligned} &T(\lambda_{0};x)\geq\tau_{T},\\ &d_{\mathbb{S}^{1}}(\phi(\lambda_{0};x),\phi^{\star})\leq\varepsilon_{\phi}\end{aligned}$

Appendix

Table S1: Abstract task-family taxonomy used to describe the benchmark in this work.

Setting	Split	Tasks	Family coverage	Interpretation
Same-type	Train	50	One sibling template from each of $G_{1}$ through $G_{6}$	Same-type training over all six primary families
Same-type	Validation	15	Held-out sibling of $G_{1}$ , plus $G_{\mathrm{aux}}$	Mild generalization check using closely related sibling templates
Same-type	Test	50	Held-out sibling templates from $G_{2}$ through $G_{6}$	Same-type test on sibling templates not used during training
New-type-A	Train	50	$G_{1}$ , $G_{2}$ , and $G_{3}$	Training restricted to single-condition, dual-angle, and polarization-selective reflection
New-type-A	Validation	15	$G_{4}$	Validation on an unseen dual-wavelength family
New-type-A	Test	50	$G_{5}$ , $G_{6}$ , and $G_{\mathrm{aux}}$	Family-heldout test across plasmonic and transmissive phase-target regimes

Appendix

Table S2: Abstract composition of the finalized splits for the three evaluation settings.

	Same-type		New-type-A		New-type-B
Row	SG (%)	CPF	SG (%)	CPF	SG (%)	CPF
Starter-skill baseline (test)	38	0.51	92	0.96	20	0.34
Iteration 1 (train / val)	48 / 100	0.58 / 1.00	52 / 100	0.61 / 1.00	92 / 13	0.96 / 0.23
Iteration 2 (train / val)	80 / 100	0.89 / 1.00	48 / 93	0.62 / 0.93	92 / 100	0.96 / 1.00
Iteration 3 (train / val)	80 / 100	0.90 / 1.00	42 / 80	0.44 / 0.80	92 / 100	0.96 / 1.00
Iteration 4 (train / val)	80 / 100	0.87 / 1.00	26 / 100	0.34 / 1.00	90 / 100	0.95 / 1.00

Appendix

Table S3: Training and validation dynamics across skill-evolution iterations, for the three settings.

Figure S5.1: Tool-call composition during skill evolution, by setting (columns: Same-type, New-type-A, New-type-B) and split (top row: training; bottom row: validation). Stacked bars give the per-task share of read, bash, and write tool calls at each iteration; the black line is the total calls per task (right axis) and the dashed line is the fixed starter-skill baseline. Read and bash calls correspond to exploration and write calls to revision, so a shift toward shorter, write-led cycles indicates less effort spent on broad search as the skill matures.

Category	Definition
Tensor Index OOB	Errors caused by invalid tensor indexing, shape mismatch, or out-of-bounds tensor access during candidate-program execution.
API Misuse	Errors caused by incorrect use of solver or optimization APIs, including invalid argument conventions and unsupported call patterns.
Gradient Error	Failures related to broken computational graphs, missing gradients, or invalid backward propagation through the optimization objective.
No Code / No Solution	Cases in which the Coding Agent fails to produce a complete executable candidate, or produces a candidate that never reaches a valid solution.
Excluded infrastructure failures	Connection errors, timeouts, and similar runtime-environment failures are excluded from Figure 3 because they are not directly attributable to the learned skill content.

Appendix

Table S4: Error categories used in the main error-analysis figure.

Snapshot	Source	Length	Primary role
starter skill	same-type run, starter-skill snapshot (iter0)	92 lines	Minimal TorchRDIT setup and execution pattern
Evolved skill	same-type run, archived selected snapshot (iter1)	218 lines	Structured task classification, failure-aware guidance, and reusable strategy selection

Appendix

Table S5: Representative skill snapshots used in the qualitative case study.

Explore similar work

Aug 26, 2026

Towards a universal meta-optics solver via large language models

Metasurface design increasingly requires fast models that can operate across structurally distinct device families, rather than retraining a separate surrogate for every geometry class. Conventional neural network surrogates often depend on fixed-dimensional descriptors, family-specific output formats, and repeated architecture tuning, which limits their scalability across heterogeneous meta-atoms. Here, we present a unified large language model (LLM) workflow for multi-family metasurface modeling and inverse-design. Geometries, design parameters, and optical response channels were converted into a shared instruction-following text format and used to fine-tune Gemma-2-9B across 8 metasurface families. Compared with single-family baselines, the joint model simultaneously predicted the optical responses of all metasurface families while reducing the MSE for each family by an average of 56.5%. The same representation was also used for inverse design. These results show that a shared sequence-based LLM interface can provide a practical route to cross-family metasurface design while reducing the need for task-specific surrogate architectures.

Huanshu Zhang, Lei Kang, Yuyan Chen +3

The Pennsylvania State University, Department of Electrical Engineering, University Park, PA 16802, USA · School of Electronics and Information Technology, Sun Yat-Sen University, Guangzhou 510275, China

Neural Surrogate Modeling Surrogate Modeling Inverse Design

Ask PDF

Oct 1, 2026

AbsorbEvo: An Agentic Framework for Autonomous Inverse Design of Microwave Absorbers

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.

Zhicheng Feng, Yubo Zhao, Xuefeng Yao

Department of Engineering Mechanics, Applied Mechanics Lab, Tsinghua University, Beijing 100084, P. R. China

Large Language Model-Guided Optimization Natural Language to Optimization Modeling Inverse Design Autonomous Experimentation

Ask PDF

Jul 6, 2026

MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution

Recent LLM agents tackle increasingly long-horizon, open-ended tasks, and external skills, reusable procedural knowledge supplied to the agent, further extend this capability. However, a fixed, hand-authored skill is rarely optimal, and cannot adapt to the diversity of tasks an agent encounters. Self-improving agents address this by rewriting their own skill files from execution traces, yielding meaningful gains on challenging benchmarks. Yet such self-evolution remains non-recursive: it improves only the task skill (what the agent does) while the improvement procedure (how it improves) is authored once and held fixed. We introduce MetaSkill-Evolve, a two-timescale framework that makes agentic skill improvement recursive: every branch carries both a task skill

s

and a branch-local meta-skill

m=(ψ,σ,α,π,\varepsilon)

whose five components parameterise the Analyzer, Retriever, Allocator, Proposer, and Evolver agents of the improvement pipeline. Task skills evolve on a fast loop while the meta-skill evolves on a slower one under the same pipeline applied to itself, with no additional model or objective. With all five pipeline agents sharing a single frozen backbone, MetaSkill-Evolve outperforms no-skill, static-skill, and single-level evolution baselines on three agentic benchmarks (OfficeQA, SealQA, ALFWorld), improving held-out test accuracy over the raw backbone by +23.54, +16.09, and +1.92 points respectively.

Zefeng Wang, Minxi Yan, Jinhe Bi +3

LMU Munich · The Chinese University of Hong Kong · MCML +1

LLM Agent Self-Improvement LLM Agent Skill Learning Recursive Self-Improvement

Ask PDF