PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhicheng Feng, Yubo Zhao, Xuefeng Yao

arXiv:2610.01119v1cs.AI

TL;DR

High-performance microwave absorber optimization requires joint material and geometry design with costly full-wave evaluation. AbsorbEvo maps natural-language objectives to candidates screened by a physics prior and verified with full-wave simulations; on held-out tasks it achieved 79.17% task success and 0.7816 mean best coverage. The results support a methodological basis for natural-language-driven autonomous absorber design, while validation beyond MAHSS and TPMS remains needed.

  • Problem

    Jointly optimizing absorber materials and geometry through high-fidelity full-wave evaluation remains cumbersome and time-consuming.

    Source: Introduction · p. 2

  • Method

    AbsorbEvo translates natural-language objectives into absorber designs, using language-guided proposals and a physics prior to screen candidates before full-wave evaluation.

    Source: Introduction · p. 4

  • Results

    79.17% task success and 0.7816 mean best coverage were achieved across 24 held-out tasks with five proposal opportunities, exceeding Generic Agent and No-feedback on both measures.

    Source: Discussion · p. 15

  • Takeaways & Limitations

    The work provides a methodological basis for natural-language-driven autonomous absorber design.

    Source: Discussion · p. 16

    Evaluation used MAHSS and TPMS as representative geometries; other structures, materials, and incidence conditions require further validation.

    Source: Discussion · p. 16

Abstract

from arXiv · show

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.

Introduction

This section introduces AbsorbEvo, which translates natural-language objectives into full-wave-verified absorber designs through language-guided search, physics-based feedback, and independently validated textual skills. It also reports full-wave task results and examines skill reuse, physics-prior ranking, and individual framework components.

Introduction

  • Joint material–geometry optimization remains cumbersome because absorber performance couples material properties, geometry, and incidence conditions, while repeated full-wave simulations are costly.Expertise is also needed to select models, configure tools, and adjust search strategies.Source: Introduction · p. 2
  • AbsorbEvo translates natural-language performance objectives into absorber designs through candidate evolution guided by language reasoning and physics-based prediction.Its language model uses task requirements and search history to propose adjustment directions and magnitudes; experience from training tasks is distilled into textual skills and independently validated before reuse.Source: Introduction · p. 4
  • AbsorbBench-36 spans explicit and implicit geometries, material combinations, and incidence conditions to assess target attainment and search efficiency.The study also analyzes optimization trajectories and electromagnetic responses and tests cross-task skill transfer.Source: Introduction · p. 4

Overview of the AbsorbEvo framework

  • Each fixed-budget round combines 36 directed candidates with 12 global candidates, ranks them using a low-cost physics prior, and submits only the top-ranked design to full-wave simulation.Results passing validity checks enter task history to guide subsequent search.Source: Overview of the AbsorbEvo framework · p. 5
  • Design quality is measured as the target-band fraction meeting the reflection-loss threshold, averaged equally across task-defined incidence conditions.Full-wave results, rather than theoretical predictions, determine target attainment and the final best design.Source: Overview of the AbsorbEvo framework · pp. 5–6
  • The framework distils training-task actions, computational results and error records into textual skills, independently validates them, and keeps this external guidance fixed during held-out evaluation.The reported evaluation distinguishes task execution, within-task design improvement and cross-task experience reuse.Source: Overview of the AbsorbEvo framework · pp. 6–7

Physics-prior agreement and system validation

  • Prior predictions are used to screen parameters, but design updates and target-attainment assessments remain based on full-wave evaluation because selected-design agreement does not guarantee accuracy for every candidate.Correlation measures rank agreement, whereas absolute errors quantify deviations in predicted coverage.Source: Physics-prior agreement and system… · p. 7
  • B2 and B3 met targets with valid results on all seven full-wave tasks, while B0 and B1 each failed physical validity checks on one task.All four configurations completed simulations; passing physical validity checks did not always imply target attainment.Source: Physics-prior agreement and system… · pp. 7–8, Figure 3 · p. 8
  • Predictive agreement was assessed by paired comparisons of theoretical predictions with existing full-wave results for selected TPMS and MAHSS designs, using identical geometric and material configurations, target frequency bands, reflection-loss thresholds, and incidence conditions.Source: Physics-prior agreement and system… · p. 7
  • The evaluator detected all 20 predefined faults, including complex-power and Floquet-mode-mapping errors, with no false positives among 17 fault-free controls.These outcomes demonstrate fault detection within the tested set.Source: Figure 3 · p. 8, Physics-prior agreement and system… · p. 8
  • In five ablation tasks, the full configuration succeeded on all five; removing range filtering, removing physics diagnosis, or reverting to the initial skill reduced successes to three, four, and three.Without runtime evaluator feedback, the configuration still completed all five tasks.Source: Figure 3 · p. 8, Physics-prior agreement and system… · p. 8

Target attainment and proposal efficiency under a fixed budget

The section examines AbsorbEvo’s task success and best coverage in fixed-budget held-out comparisons, alongside cases illustrating how screening, full-wave feedback, and validated skill transfer supported design search.

Target attainment and proposal efficiency under a fixed budget

  • 19 of 24 tasks (79.17%) succeeded with AbsorbEvo, versus six (25.00%) for Generic Agent, 15 (62.50%) for No-feedback, and three (12.50%) for Random.All methods received the same initial design and five proposal opportunities per task.Source: Target attainment and proposal… · p. 9, Figure 4 · p. 10, Table 2 · p. 11
  • By the third round, AbsorbEvo had succeeded on 14 tasks, versus Generic’s two then and six after all five rounds.On the five TPMS tasks where both succeeded, mean first-success round was 2.60 for AbsorbEvo and 3.80 for Generic.Source: Target attainment and proposal… · p. 9, Figure 4 · p. 10
  • Across structural families, AbsorbEvo succeeded on 9/12 MAHSS tasks, versus none for Generic or Random; on TPMS, AbsorbEvo and Generic succeeded on 10/12 and 6/12 tasks, respectively.Source: Target attainment and proposal… · p. 9
  • AbsorbEvo achieved mean best coverage of 0.7816, versus 0.6434 for Generic, 0.6448 for Random, and 0.7613 for No-feedback.Across 120 proposal opportunities per method, AbsorbEvo produced 53 target-meeting designs (44.17%), versus 13 (10.83%) for Generic, five (4.17%) for Random, and 28 (23.33%) for No-feedback.Source: Table 2 · p. 11, Target attainment and proposal… · p. 11

Physics-guided search and design evolution

  • In an independent two-dimensional MAHSS case with material parameters fixed and coating thickness and core height varied, the first-round global candidate increased coverage; only the top-ranked candidate underwent full-wave evaluation.Source: Physics-guided search and design… · p. 11

External skill evolution and transfer

  • On eight held-out tasks, validated skill C1 succeeded on seven, compared with five for starter S0, with both additional successes in the out-of-distribution group.On six validation tasks, C1 succeeded on all six versus four for S0; C1 was selected through independent validation rather than model self-assessment.Source: External skill evolution and transfer · pp. 12–13, Figure 6 · p. 13

Representative physics-verified design cases

  • For TPMS case AB36-TPMS-T07, coverage increased from 45.24% to 90.95% in round two, yielding the best design across five rounds.The design retained the Diamond topology but used M6 instead of the initial material.Source: Figure 7 · p. 14, Representative physics-verified design… · p. 15
  • For M4, solid volume fraction rose from 0.29 to 0.44 and period from 15.4130 to 19.7074 mm; longest contiguous qualifying bandwidth rose from 4.524 to 9.095 GHz.All three subsequent rounds still met the target, but coverage remained 90.79%, below the second-round result.Source: Representative physics-verified design… · p. 15

Discussion

  • Across 24 held-out tasks with five proposal opportunities each, AbsorbEvo achieved 79.17% task success and 0.7816 mean best coverage, exceeding Generic Agent and No-feedback on both measures.Source: Discussion · p. 15
  • Candidate screening and full-wave feedback played complementary roles: No-feedback retained candidate generation and prior screening, solved more tasks than Generic Agent and Random, while new full-wave feedback enabled additional solutions.Source: Discussion · p. 16
  • Independent physics evaluation distinguished valid computational results from designs meeting performance targets, so prior predictions supported candidate screening but did not replace full-wave verification.Source: Discussion · p. 16
  • Validated textual skills improved success on new tasks while preserving all previously successful outcomes; skill content and model weights stayed fixed during held-out testing.This outer-loop transfer used previously selected guidance without updates driven by held-out results or changes to base-model weights.Source: Discussion · p. 16
  • Evaluation covered MAHSS and TPMS as explicit- and implicit-geometry representatives; other structures, materials, incidence conditions, and alternative predictors require further assessment.Source: Discussion · p. 16

Natural-language task interpretation and automated modelling

This section describes how AbsorbEvo translates natural-language absorber objectives into simulation-ready designs and uses structured task and geometry representations, adaptable design-space models, and explicit physical checks to guide candidate search. It covers physics-based priors, including effective anisotropic models for predicting reflection spectra, and full-wave simulations for final performance evaluation.

Natural-language task interpretation and automated modelling

  • The planning agent converts natural-language requirements into an AbsorberSpec and SimulationRequest covering design variables, bounds, target bands, incidence conditions, geometry, materials, and solver settings.Explicit models define structures through dimensions and component relationships, while implicit models use spatial functions and isosurfaces; both connect through a unified modelling interface.Source: Natural-language task interpretation… · pp. 16–17

Deterministic physics verification

  • Only full-wave results passing independent checks of geometry, materials, excitation, output completeness, required modes, power relations, and solver status count as valid performance evidence.Coverage sums all qualifying bandwidth intervals within the target band and averages across task-specified incidence conditions; both adjacent sampling points must meet the reflection-loss threshold for an interval to count.Source: Deterministic physics verification · pp. 17–18

Adaptive multi-fidelity optimization

  • The agent proposes directed parameter changes from task objectives and search history, adds global samples, ranks candidates with a physics prior, and simulates the top-ranked design each round.Valid full-wave results inform subsequent adjustments; the candidate-generation reference is the valid new proposal with the highest coverage, or the shared initial design if none is available.Source: Adaptive multi-fidelity optimization · p. 19

Design spaces and physics priors

  • The screening prior uses effective-medium theory and layered-media wave propagation, and the framework can also use pretrained, validated surrogate models.Design spaces include explicit MAHSS structures and implicit TPMS structures; the same material parameters are used in prior predictions and full-wave simulations.Source: Design spaces and physics priors · p. 20
  • MAHSS models an effective anisotropic core within the full layer stack, whereas TPMS uses a single effective anisotropic layer; both use PEC backing.Source: Design spaces and physics priors · pp. 20–21
  • Wave-propagation relations produce predicted reflection spectra for both structural families.For MAHSS, input impedance is calculated through the layer stack; the TPMS model calculates reflection from its single effective layer.Source: Design spaces and physics priors · p. 21
  • Target bands, thresholds, and incidence conditions determine CTheory for design comparison, while design updates and final evaluation use full-wave CExact.Source: Design spaces and physics priors · p. 21

Experience accumulation and external skill evolution

The section describes how computational experience is distilled into validated textual skills, and lays out system, fixed-budget, and transfer evaluation protocols. It also identifies benchmark and code availability.

Experience accumulation and external skill evolution

  • Experience from training-task actions, valid responses, and errors is distilled into textual rules for modelling, tool use, result checking, and parameter adjustment.Successful outcomes preserve effective practices, while unmet targets and computational errors highlight areas for refinement.Source: Experience accumulation and external… · p. 22
  • Candidate skills replace the current version only when validation performance improves without deterioration in physical validity, numerical quality, or error diagnosis.Target attainment, task success, performance metrics, and computational cost are compared in a predefined priority order; ties retain the existing skill.Source: Experience accumulation and external… · p. 22

System-level evaluation

  • System evaluation covers 24 tasks, including seven full-wave tasks and an offline error-diagnosis task, with 20 predefined faults and 17 fault-free controls.The full-wave tasks separately record execution, convergence, physical validity, and target attainment; ablations use five independent tasks and are reported separately from the fixed-budget comparison.Source: System-level evaluation · p. 22

Fixed-budget optimization

  • AbsorbBench-36 contains 18 MAHSS and 18 TPMS tasks, and the fixed-budget comparison gives each method five proposal opportunities per task.The 24 held-out comparison uses shared initial designs and their full-wave results; Random, Generic Agent, AbsorbEvo, and No-feedback are compared under the stated setup.Source: Fixed-budget optimization · p. 23
  • AbsorbEvo proposes adjustments, generates candidates, and uses a physics prior to select a design for full-wave evaluation; No-feedback omits new full-wave feedback from subsequent adjustments.No-feedback retains the initial design as its reference, while full-wave results are used only for recording and scoring.Source: Fixed-budget optimization · p. 23
  • Success requires a constraint-compliant candidate, valid full-wave results, and coverage meeting the task target within five rounds.Best valid coverage and first-success round are also used to assess design quality and search efficiency.Source: Fixed-budget optimization · p. 24

Skill evolution and transfer

  • For transfer evaluation, an evolved skill is selected on six validation tasks, then compared with the initial skill on the same eight held-out tasks without test-time updates.The held-out tasks span same-type, cross-family, and out-of-distribution categories; failure-to-success and success-to-failure transitions are recorded.Source: Skill evolution and transfer · p. 25

Data availability

Funding

The funding statement names support from a National Science and Technology Major Project of China. The authors report no known competing financial interests or personal relationships that could have influenced the work.

Funding

  • The work was supported by the National Science and Technology Major Project of China.The provided statement begins to identify a grant but does not provide its number.Source: Funding · p. 27

Competing interests

  • The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.Source: Competing interests · p. 27

Supplementary Information

The supplementary information documents the structures, simulation settings, physics-prior implementations, physical verification criteria, and information provided to competing methods. It also specifies information-access limits and the fixed-budget rules for candidate generation, ranking, and selection.

S1.2. Implementations of the physics priors

  • The MAHSS physics prior models the coated honeycomb core as an effective diagonal-permittivity layer and recursively calculates five-layer input impedance from the PEC boundary.It uses area-weighted end-wall and inclined-wall regions and different tensor components for TE and TM propagation.Source: S1.2. Implementations of the physics… · p. 32
  • The TPMS prior derives effective properties from solid fractions in equal-thickness slices; identical material, thickness and mean solid fraction yield the same approximate response.The prior neglects cross-polarized coupling, whereas full-wave evaluation includes co- and cross-polarized reflected power.Source: S1.2. Implementations of the physics… · p. 32
  • The historical comparison included 474 MAHSS cases and 474 of 531 eligible TPMS records, with Prior and Exact evaluated on identical task conditions and frequency grids.Cases were deduplicated by geometry, material, incidence angle, target band and threshold; no additional full-wave simulations were run, and agreement metrics apply only to selected cases.Source: S1.2. Implementations of the physics… · pp. 32–33

S1.3. Physical verification criteria

  • Physical validity is assessed separately from target attainment, requiring complete finite TE00/TM00 co- and cross-polarized spectra and power-consistency residuals below 10^-10.The cross-channel complex-amplitude difference threshold is 10^-4 and captures magnitude and phase differences.Source: S1.3. Physical verification criteria · p. 33, Table 3 · p. 34
  • In the fixed-budget comparison, energy and power residuals were not independently checked for every result and do not constitute field-integration energy closure.Geometry and material checks were limited to build records, and checking the two fundamental modes did not verify all higher-order propagating modes.Source: S1.3. Physical verification criteria · pp. 33–34

S2.1. Information access

  • The four methods differ in candidate generation and access to physics-prior and full-wave feedback: AbsorbEvo uses both, while No-feedback omits new Exact results.Random samples designs randomly, and Generic Agent directly proposes parameters; AbsorbEvo and No-feedback share fixed domain guidance, unlike Generic Agent.Source: S2.1. Information access · p. 34
  • AbsorbEvo uses current-task records without access to solutions from other tasks or methods, while No-feedback receives no new Exact responses during current optimization.Source: S2.1. Information access · p. 35

S2.2. Candidate generation and selection

  • Each formal-comparison round generates 36 directed and 12 global candidates, with LLM-chosen variable directions, normalized steps, and overall scales under fixed candidate and round counts.At least two variables must be enabled, and at least one must have a nonzero adjustment direction.Source: S2.2. Candidate generation and selection · p. 35
  • Directed search spans single-, two-, and multivariable adjustments; remaining local slots use Halton sampling, while disabled variables retain their values.Continuous variables are normalized to their design ranges and categorical variables mapped to equally spaced indices; single-variable multipliers are 0.500, 1, 1.500, and 2.Source: S2.2. Candidate generation and selection · p. 35
  • Global candidates sample the full design domain independently of the LLM plan, and out-of-range or duplicate candidates are discarded.AbsorbEvo and No-feedback use identical seeds for corresponding tasks and rounds; duplicate checks include the current pool, reference point, and previous pools.Source: S2.2. Candidate generation and selection · p. 36
  • After the complete candidate pool is formed, prior-ranked candidates are ordered by prediction validity, predicted coverage, and EAB, and only the top design is submitted to Exact each round.Tie-breakers include response shortfall, normalized distance from the generation reference, and candidate index.Source: S2.2. Candidate generation and selection · p. 36
  • AbsorbEvo generates from the valid new proposal with highest coverage and selects the final best design from the initial design and valid proposals, retaining the earlier design in ties.No-feedback uses the shared initial design as its generation reference.Source: S2.2. Candidate generation and selection · p. 36

S2.3. Proposal budget and retries

Each task has a five-opportunity proposal budget, with success requiring a new proposal to pass validity and task checks and meet its target. An independent two-variable MAHSS example illustrates proposal outcomes and verified-best selection.

S2.3. Proposal budget and retries

  • Each task receives five proposal opportunities; success requires at least one new proposal to pass physical-validity and task-constraint checks and meet the target.If none succeeds, success is recorded as not observed within the five-round budget, and the final best coverage may still come from Initial.Source: S2.3. Proposal budget and retries · pp. 37–38
  • An incomplete candidate pool still consumes an opportunity, receives no replacement proposal, and has its coverage recorded as missing.The incomplete pool does not proceed to Exact.Source: S2.3. Proposal budget and retries · p. 37
  • Each selected candidate is evaluated separately at each specified angle, with at most 10 attempts per candidate–angle pair and unchanged geometry, materials, and conditions during retries.Interrupted executions are resumed or checked before outcomes are determined, and all five rounds are completed even if the target is reached earlier.Source: S2.3. Proposal budget and retries · pp. 37–38

S2.4. Independent two-variable illustration

  • The independent example fixes Debye parameters, varies only b (0.005–0.302 mm) and h (3–19.200 mm), and targets 0.950 coverage from an initial 0.481 over 2–18 GHz at 0° TE.The target uses an absorptance threshold of 0.900.Source: S2.4. Independent two-variable… · p. 38
  • P2 first meets the target, P4 is best among the five proposals, and the dashed branch marks P5 in the verified-best trajectory.Grey diamonds mark pre-validation points withheld from the agent.Source: S2.4. Independent two-variable… · p. 40

S2.5. System-level evaluation configurations

The system-level evaluation compares configurations across held-out tasks and evaluates ablations separately. All configurations use the same evaluator, while the ablation tests retain final evaluation even when runtime feedback is removed.

S2.5. System-level evaluation configurations

S3.1. Tasks and data splits

AbsorbBench-36 contains MAHSS and TPMS tasks across development, validation, and held-out splits. The formal comparison uses held-out tasks with predefined objectives, design domains, and initial designs.

S3.1. Tasks and data splits

  • AbsorbBench-36 comprises 18 MAHSS and 18 TPMS tasks, divided into 8 development, 4 validation, and 24 held-out tasks.The formal comparison uses the 24 held-out tasks, with 12 from each structural family.Source: S3.1. Tasks and data splits · p. 42
  • Task objectives specify target bands, incidence conditions, reflection-loss thresholds, and required coverage; design domains and initial designs are listed separately.The supplied TPMS objectives include both single-angle and multiple-angle conditions, with TE, TM, or both polarizations.Source: S3.1. Tasks and data splits · pp. 42–43, Table 6 · p. 42
  • Task objectives, design ranges, shared initial designs, and data splits were fixed before the formal comparison.Source: S3.1. Tasks and data splits · p. 42
  • For PEC-backed structures, RL thresholds of −5 dB and −10 dB correspond to absorptance thresholds of 0.684 and 0.900, respectively.Source: S3.1. Tasks and data splits · p. 43

S3.2. Design domains and initial designs

Supplementary Table 7 specifies task-specific design domains and initial designs for MAHSS and TPMS tasks. Only listed variables are adjusted; all other conditions remain fixed.

Design domains and initial designs

MAHSS design domains

  • MAHSS tasks vary coating thickness, core height, and Debye parameters within task-specific ranges, while cell geometry settings can remain fixed.For example, AB20-HC-T01 lists ranges for b, h, ε∞, εs, and τ, with cell angle 60°, s = 2.550 mm, and w = 0.055 mm.Source: S3.2. Design domains and initial designs · pp. 44, 47

Initial designs

TPMS design domains

  • TPMS task domains vary level half-width or target solid volume fraction and period, with topology and material choices specified by task.TPMS tasks use a 64³ grid, one period per axis, xyz orientation, and PEC backing; VF is mapped to t as described in S1.2.Source: S3.2. Design domains and initial designs · pp. 46–47

S4.1. Training, validation and transfer tasks

The supplement specifies training, validation, and held-out transfer tasks, then describes skill selection and transfer-cost accounting. In the held-out comparison, S0 and C1 each completed eight attempts at four full-wave-equivalent cost units.

S4.1. Training, validation and transfer tasks

  • The independent skill evaluation includes 10 training, 6 validation, and 8 held-out tasks, separate from the 24 AbsorbBench-36 optimization tasks.Held-out tasks cover same-type, cross-family, and out-of-distribution conditions.Source: S4.1. Training, validation and… · p. 48
  • Tasks use 2–18 GHz normal-incidence TE/TM with PEC backing, and success requires executable, physics-valid, converged status plus task-specific thresholds.Mean and Peak are the minimum of the corresponding TE and TM statistics; EAB and an additional relative-gain condition are excluded from the success score.Source: S4.1. Training, validation and… · p. 49

S4.2. Skill content and selection

  • Skill evolution distills design actions, outcomes, errors, performance, and cost into textual guidance, preferring shorter text after these criteria.C1 adds material-reference checks, Floquet-mode prechecks, and guidance for single-variable coating adjustments relative to S0.Source: S4.2. Skill content and selection · p. 50
  • C1 was selected after the first validation round and was not replaced by later candidates.Candidates are compared on the same six validation tasks, with physical validity, convergence, and diagnosis protected from decline.Source: S4.2. Skill content and selection · p. 50

S4.3. Transfer evaluation and cost accounting

  • Transfer cost counts eligible Improve/Transfer task evaluations, including identical-input results reused during testing, while excluding pre-solve failures and offline Debug tasks.The cost is independent of target attainment and is intended for comparable task-level accounting.Source: S4.3. Transfer evaluation and cost… · p. 51
  • S0 and C1 each completed 8 held-out task attempts and incurred 4 full-wave-equivalent cost units.This accounting is task-level cost, not actual solver calls or elapsed time.Source: S4.3. Transfer evaluation and cost… · p. 51
Loading 2610.01119v1…