PCFBench: How Far Can Large Vision-Language Models Go in Physics-aware Photonic Inverse Design?

Shengchao Chen; Ting Shu; Sufen Ren

cs.CVOct 23, 2026

PCFBench: How Far Can Large Vision-Language Models Go in Physics-aware Photonic Inverse Design?

Authors: Shengchao Chen, Ting Shu, Sufen Ren

Organizations: Australian AI Institute, University of Technology Sydney · School of Artificial Intelligence, Shenzhen University · School of Information and Communication Engineering, Hainan University

Abstract

Photonic crystal fibers (PCFs) underpin critical applications in telecommunications, sensing, and high-power laser delivery, yet their design remains a manual, expert-intensive process that couples geometric intuition with deep waveguide physics. Vision-language models (VLMs) offer a promising path toward automating PCF design, but can they truly reason about photonic physics from microscopy images, or do they merely recognize visual patterns? We introduce PCFBench, the first comprehensive benchmark for evaluating VLMs on the perception, reasoning, and decision steps of the photonic inverse-design loop. PCFBench spans 26 tasks organized into six categories of increasing cognitive demand (Geometry Perception, Physics Understanding, Text Generation, Multi-modal Reasoning, Inverse-Design Decision, and Code Generation), built from over 220K FDTD simulation samples across 18 structurally diverse fiber families. We evaluate both individual VLMs on isolated tasks and multi-agent systems (MAS) on the full sequential pipeline from perception to design. Our results reveal a striking perception-design gap: current models can identify fiber geometry with moderate success, yet consistently fail to translate that visual understanding into physics-aware design actions. Moreover, cascading agents through the pipeline exposes significant error propagation, where upstream perception mistakes compound into downstream design failures. These findings suggest that neither model scaling nor naive agent chaining suffices, and that closing the loop demands explicit physical grounding. PCFBench provides the community with a rigorous, multi-dimensional testbed to measure progress toward AI-assisted photonic engineering. The project is available at https://pcfbench.github.io/.

Figures & tables

Figure 1: PCFBench overview. PCFBench formulates PCF inverse design as a multi-stage, multimodal reasoning problem, where complementary visual inputs, geometric cross-sections and mode fields, are processed through a sequential pipeline of perception , understanding , reasoning , and design . The benchmark comprises 26 tasks organized into a six-dimensional hierarchy spanning geometry perception, physics understanding, text generation, multi-modal reasoning, inverse-design decisions, and code generation, arranged along an increasing difficulty gradient that mirrors real-world workflows. Built on over 220K multimodal samples with human expert annotations across 18 fiber families, PCFBench supports both single-model and multi-agent evaluation.

Figure 2: PCFBench is constructed through a unified pipeline: ① structure extraction from diverse fiber geometries, ② FDTD simulation of multi-wavelength mode fields, ③ aggregation of geometric and physical metadata, ④ multi-level expert annotation, ⑤ expert quality control via re-simulation, and ⑥ a multi-dimensional evaluation suite for single-model and multi-agent assessment.

Figure 3: Dataset composition of PCFBench. The 220K PCF samples span 18 canonical families. Each label gives the family’s sample count and proportion.

Figure 4: Representative structures (left) and corresponding mode fields at $\lambda=1550$ nm (right) for all families. The paired views reveal distinct guidance mechanisms: hollow-core families exhibit dark cores with bright anti-resonant rings, solid-core families show compact guided modes, and specialty geometries (plasmonic, vortex, multi-core) produce characteristic field patterns. See Appendix A.2 for additional exemplars per family.

Hierarchy	#	Task	Category	Img	Metric
Percep.	1	Family Classif.	Geo. Percep.	S	Acc. $\uparrow$
	2	Param. Pred. (Pitch)	Geo. Percep.	S	NRMSE $\downarrow$
	3	Param. Pred. (Hole D)	Geo. Percep.	S	NRMSE $\downarrow$
	4	Structure Counting	Geo. Percep.	S	Exact Acc. $\uparrow$
	5	Topology Recog.	Geo. Percep.	S	Acc. $\uparrow$
Underst.	6	Loss Pred. (1310 nm)	Phys. Und.	F 1310	Log-MAE $\downarrow$

Table 1: Overview of the 26 PCFBench tasks (hierarchy in Fig. 1 ). “S” = structure image, “F” = mode-field image.

Panel of Table 2 — Table 2: PCFBench leaderboard: one representative metric per task across 6 dimensions (full results in Tabs. 16 – 18 ). Top-3 per column are Gy Gold / Gy Blue / Gy Gray .

			GPT-4o-mini			Qwen-VL-Max			Gemini-3.1-Lite
Method	Coord.	#Ag.	PSel	CL	HS	PSel	CL	HS	PSel	CL	HS
Vanilla	Single	1	.110	.540	.34	.360	.580	.45	.240	.270	.28
CoT	Single	1	.110	.490	.32	.440	.520	.45	.240	.270	.28
Self-Consistency	Voting	$N$	.110	.590	.36	.410	.520	.44	.230	.270	.28
LLM Debate	Debate	3	.120	.500	.33	.330	.530	.42	.170	.290	.26
MAD	Debate	3+	.290	.731	.48	.126	.411	.30	.280	.462	.37

Table 3: Multi-agent inverse design results. Metrics: PSel (Param. Selection Acc.), CL (Closed-Loop Success), and HS (Human Score). #Ag. is the number of agents per run, and

N,M

are method counts (Tab. 24 ). Sequential Pipeline reports Cascaded and Oracle (perfect upstream) modes.

Figure 5: Scaling curves: parameter count vs. category-averaged performance.

	Family Acc				CL Succ.
Model	$k{=}0$	$k{=}2$	$k{=}4$	$k{=}8$	$k{=}0$	$k{=}2$	$k{=}4$	$k{=}8$
Gemini-3-Flash	.668	.706	.741	.763	.680	.682	.676	.681
GPT-5	.455	.521	.566	.592	.620	.626	.617	.624
GPT-4o-mini	.130	.217	.259	.291	.800	.794	.791	.796
Qwen2.5-VL-72B	.185	.224	.251	.268	.589	.593	.584	.590
Qwen2.5-VL-7B	.096	.134	.159	.172	.638	.643	.634	.641

Table 4: Family accuracy and closed-loop success with

k

in-context exemplars (mean over five support-set draws).

Figure 7: Left Cross-task Spearman $\rho$ . Right Global ranking (mean normalized rank, 19 metrics).

Table 12

Figure 8: Closed-loop inverse design on two PCFs (left: ARF, right: dual-core coupler). Top: structures with $\Delta$ -loss, bottom: mode fields and ranking (negative = improvement). More cases in Appendix D.6 .

Appendix figures & tables42 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: The 220,000-sample corpus, split into four groups by guidance mechanism and then into 18 families. Node heights and ribbon widths are proportional to sample counts.

Figure 10: Dataset distributions. (a–b) Pitch and hole diameter with KDE (dashed: median). (c) Ring count. (d–e) Log-scale loss at 1310 nm and 1550 nm. (f) Annotation-level coverage.

Figure 11: Three exemplars per family from the 220,000-sample corpus. Each structure (upper row) sits above its FDTD-simulated fundamental mode field at $\lambda=1550$ nm (lower row).

Check	Rejection Criterion
Numerical convergence	Confinement loss is undefined or infinite.
Mode-pattern validity	The mode-field image shows isotropic speckle without a discernible confined-mode lobe.
Geometric validity	Wall thickness $\leq 0$ , overlapping holes, core radius $>$ cladding outer radius, or strut count outside the family range.
Mode-count consistency	The estimated mode count deviates from the $V$ -number prediction by more than one order.
Annotation–physics consistency	The L3 diagnosis proposes a parameter change whose direction contradicts the family-specific physics rule (Appendix D.5 ) for the L2-reported loss regime.

Appendix

Table 8: Automated quality-control checks applied to every candidate sample. A sample is rejected when it meets the listed criterion.

Benchmark	Domain	Modality	Tasks	Samples	Closed-Loop Sim.	Inverse Design	Native MAS Eval
MathVista [ 8 ]	Math Reasoning	I + T	6 (subtypes)	6,141	–	–	–
MMStar [ 7 ]	General Multimodal	I + T	6 (subtypes)	1,500	–	–	–
GeoBench [ 16 ]	Geospatial Reasoning	I + T	12	2,000	–	–	–
PICBench [ 26 ]	Photonic IC Layout	T + C	4	420	✓ (partial)	✓	–
ScienceAgentBench [ 12 ]	Scientific Workflows	T + C + I	102 (workflows)	102	✓ (per task)	–	✓
NewtonBench [ 14 ]	Physical-Law Inference	T + I	12	1,200	–	–	–

Appendix

Table 9: PCFBench vs. prior multimodal and scientific benchmarks. Modality : I = image, T = text, C = code, D = dialogue. Closed-Loop Sim. : ground truth verified by re-running a numerical solver on the model output. Inverse Design : tasks ask for actionable parameter changes, not only classification or QA. Native MAS Eval : built-in evaluation of multi-agent systems.

Component	Setting
Engine	Meep FDTD (time-domain, 2D)
Spatial resolution	40 pixels per $\mu$ m (uniform Yee grid, $\Delta x\approx 25$ nm)
Cell size	$50\times 50\,\mu$ m square (cell radius $25\,\mu$ m)
Boundary conditions	Perfectly Matched Layer (PML) on all four lateral edges
PML thickness	$2.0\,\mu$ m
Materials	Fused silica $n_{\mathrm{SiO_{2}}}=1.444$ (fixed across both telecom bands), air holes $n=1.000$

Appendix

Table 10: Unified simulation configuration shared by all 220K PCFBench samples. Values are held constant across the 18 fiber families so that loss and mode-field outputs are directly comparable.

Cluster	$\|\Delta\log\mathcal{L}\|$ (dex)
Index-guided	0.022
ARROW	0.038
PBG / FSM	0.051
Plasmonic	0.084

Appendix

Table 11: Per-cluster numerical-sensitivity check on the simulator. Smaller is better.

Category	Task	Max Tokens	Image Input	Output Form
Geometry Perception	Family Classification	64	[img:structure]	1 family label
	Parameter Pred. (Pitch)	128	[img:structure]	1 numeric ( $\mu$ m)
	Parameter Pred. (Hole D)	128	[img:structure]	1 numeric ( $\mu$ m)
	Structure Counting (Rings)	128	[img:structure]	1 integer
	Topology Recognition	64	[img:structure]	1 topology label
Physics Understanding	Loss Prediction (1310 nm)	128	[img:field@1310 nm]	1 numeric (dB/km)

Appendix

Table 12: Per-task decoding configuration. All tasks share temperature

T\,{=}\,0.1

and the system prompt “ You are an expert in photonic crystal fibers. ”, with a task-specific output-token budget. Multi-image rows list all images given to the model. Field-only rows are ablations of the corresponding Paired task.

Panel of Table 13 — Table 13: Evaluated model catalog. Icons match the leaderboard tables. Access : Prop. = proprietary API, Open = open-weight download. “-T” marks a thinking variant.

Dimension	Score 0.0	Score 0.5	Score 1.0
Physics Validity Does the proposed parameter change align with waveguide physics (anti-resonance theory, photonic bandgap, or index-guiding principles)?	Contradicts established physics.	Direction correct but reasoning flawed.	Direction correct and reasoning physically sound.
Quantitative Correctness Is the proposed value within a physically plausible range for the identified family (e.g., ARF strut $\in$ $0.3$ – $1.2\,\mu$ m)?	Outside known bounds.	Within bounds but misaligned with the loss regime.	Appropriate for both family and loss regime.
Fabrication Feasibility Can the proposed structure be fabricated with current stack-and-draw or CVD-based processes?	Violates minimum-feature or geometric continuity constraints.	Requires a non-standard process.	Standard fabrication route.
Performance Improvement Would the proposed change plausibly reduce confinement loss as claimed?	No improvement or makes loss worse.	Marginal improvement.	Substantial improvement ( $>$ 1 order of magnitude for high-loss, $>$ 3 $\times$ for moderate-loss).
Design Coherence Are the parameter choice, direction, and magnitude internally consistent with the diagnosed loss mechanism?	Contradictory across stages.	Partial consistency.	Fully coherent end-to-end reasoning.
Justification Quality Does the textual explanation reference the relevant physical principle (Maxwell’s equations, resonance condition, bandgap shift)?	No or incorrect justification.	Partially correct reasoning.	Explicit and precise physical reasoning.

Appendix

Table 14: Human evaluation rubric for the Inverse-Design Decision category. Each of six dimensions is scored on

\{0.0,\,0.5,\,1.0\}

by two photonics experts. The composite Human Score is their arithmetic mean across dimensions and raters.

Statistic	Value
Cohen’s $\kappa$	0.72
Weighted $\kappa$	0.68
Exact agreement	81.4%
Within-one-point agreement	96.2%
Adjudicated cases ( $>$ 1-point disagreement)	12.5%
Leaderboard Kendall $\tau$ (pre vs. post adj.)	0.97

Appendix

Table 15: Inter-rater reliability of the Human Score on the 120-sample subset.

Panel of Table 16 — Table 16: Geometry Perception and Physics Understanding. Fam. : family classification, Topo. : topology recognition, App. and Mode : application and mode classification, Mech. : mechanism reasoning. Top-3 per column are Gy Gold / Gy Blue / Gy Gray .

Panel of Table 17 — Table 17: Text Generation and Multi-modal Reasoning. Text metrics are BERTScore-F1 (BS-F1), ROUGE-L (RL), term recall (TR), and completeness (Comp.). WC, MD, and EE denote wavelength comparison, mechanism diagnosis, and evidence explanation under Paired (structure and field images) or Field -only inputs. Top-3 per column are Gy Gold / Gy Blue / Gy Gray .

Panel of Table 18 — Table 18: Inverse-Design Decision and Code Generation. P.Sel. : parameter selection, HS : Human Score (Appendix B.4 ). Code metrics are syntax correctness (Syn.), quality (Qual.), and completeness (Comp.). Top-3 per column are Gy Gold / Gy Blue / Gy Gray .

Tier	Syntax	Imports Resolve	Runs End to End	Valid Output
Best frontier	98.7%	92.4%	84.1%	79.8%
Frontier median	96.1%	86.3%	73.5%	68.9%
Best open-weight	91.8%	70.5%	55.9%	51.4%
Overall median	88.6%	64.2%	46.8%	42.5%

Appendix

Table 19: End-to-end executability of generated simulation code (pass@1, 200-sample subset).

Metric	Kendall $\tau$ (Full vs. Trimmed)	Outlier Share of Error
Log-MAE	0.97	9%
NRMSE	0.61	48%

Appendix

Table 20: Loss-metric robustness under top-1% outlier trimming.

$N$	Composite (Mean $\pm$ 95% CI)	pass@ $k$
1	$.431\pm.008$	31.2%
3	$.446\pm.005$	38.7%
5	$.450\pm.004$	41.5%
10	$.452\pm.004$	43.0%

Appendix

Table 21: Self-Consistency sample-budget sweep (

T=0.1

, GPT-4o-mini, five seeds). Composite averages the parameter-selection accuracy, direction accuracy, and closed-loop success of the voted answer. A sample counts toward pass@

k

when at least one of its

k=N

sampled answers is correct.

Model	Family Acc	Topology Acc	Dir. Pitch Acc	CL Succ.
Qwen2.5-VL-7B	.092 $\pm$ .011	.233 $\pm$ .018	.516 $\pm$ .024	.631 $\pm$ .021
Gemma-3-12B	.081 $\pm$ .014	.228 $\pm$ .019	.312 $\pm$ .026	.509 $\pm$ .022
Qwen2.5-VL-72B	.187 $\pm$ .008	.426 $\pm$ .014	.430 $\pm$ .017	.585 $\pm$ .015
Qwen3-VL-235B-A22B	.219 $\pm$ .013	.487 $\pm$ .010	.234 $\pm$ .019	.164 $\pm$ .016
GPT-4o-mini	.127 $\pm$ .009	.351 $\pm$ .012	.543 $\pm$ .015	.792 $\pm$ .012
Claude-Opus-4.5	.390 $\pm$ .008	.462 $\pm$ .012	.618 $\pm$ .010	.629 $\pm$ .009

Appendix

Table 22: Paraphrase sensitivity (mean

\pm

std over the original prompt and four rewrites, each averaged over three runs at

T=0.1

). The last row gives the mean pairwise Kendall

\tau

of model rankings.

Image Set	Fréchet Distance
COCO held-out split	1.2
ImageNet validation	28.5
PCFBench structure images	214.7
PCFBench mode-field images	342.6

Appendix

Table 23: Fréchet distance to a COCO reference set in SigLIP space (2,000 images per set).

Method	Category	Agents	Workflow
Vanilla	Single-agent	1	$A\rightarrow\text{Answer}$
CoT	Single-agent	1	$A_{\text{think}}\rightarrow\text{Answer}$
Self-Consistency	Voting	$N$	$\{A_{1},\ldots,A_{N}\}\rightarrow\text{Vote}\rightarrow\text{Answer}$
LLM Debate	Debate	2–3	$A_{1}\rightleftharpoons A_{2}\rightleftharpoons A_{3}\rightarrow\text{Answer}$
MAD	Debate	3+	$\{A_{1},\ldots,A_{N}\}\rightleftharpoons\text{Judge}\rightarrow\text{Answer}$
CAMEL	Role-playing	2+ $C$	$A_{\text{user}}\rightleftharpoons A_{\text{asst}}\xrightarrow{C}\text{Answer}$

Appendix

Table 24: Multi-agent methods evaluated on PCFBench, all with the same VLM backbone. The Workflow column shows the coordination pattern (

A

= agent,

C

= critic/verifier).

		Per-Stage ( $\uparrow$ )		Pipeline-Level
Backbone	Mode	U	R	PSel $\uparrow$	CL $\uparrow$	E2E $\uparrow$	Casc $\downarrow$	Cons $\uparrow$	$\mathrm{Acc}_{4\|1}\,\uparrow$
	Oracle	.620	.545	.485	.785	.380	—	—	—
	Cascaded	.480	.395	.325	.655	.215	.685	.580	.735
GPT-4o-mini	$\Delta$ (C $-$ O)	$-.140$	$-.150$	$-.160$	$-.130$	$-.165$	—	—	—
	Oracle	.555	.510	.580	.755	.440	—	—	—
	Cascaded	.445	.380	.475	.620	.295	.625	.615	.710

Appendix

Table 25: Diagnostic decomposition of the Sequential Pipeline. Per-stage accuracy for Understanding (U, loss-class F1) and Reasoning (R, mechanism), parameter-selection accuracy (PSel) and closed-loop success (CL) as in Tab. 3 , and the pipeline-level metrics of Appendix D.1 , under Oracle (ground-truth upstream context) and Cascaded (predicted upstream context) modes. The

\Delta

row is the Cascaded

-

Oracle gap.

Figure 12: (a) Cost-performance Pareto frontier. (b) MAS token budget vs. closed-loop success (GPT-4o-mini backbone). (c) Category profiles of the top-5 VLMs.

Family	Parameter	Direction	Rationale
Solid-Core	$\Lambda$ , $d$	increase	Stronger index guiding
	strut, gap	decrease	Reduce leakage path
ARF / Nested Anti-Resonant	strut thickness	decrease	Broaden anti-resonant band
	$\Lambda$ , core $r$	increase	Enlarge air core
	gap width	decrease	Tighter tube packing
Kagome / Suspended Cobweb	strut thickness	decrease	Fewer cladding states

Appendix

Table 26: Family-aware Design Direction rules in the high-loss regime (

\mathcal{L}>0.01

dB/km).

Cluster	Non-Monotonic	Rule Conflicts	Removed by Re-Simulation
Index-guided	4.8%	2.1%	3.6%
ARROW	8.7%	4.4%	6.9%
PBG / FSM	10.9%	5.8%	8.1%
Plasmonic	12.6%	6.3%	9.4%
Overall	—	—	5.9%

Appendix

Table 27: Non-monotonic cases, rule conflicts, and re-simulation exclusions per cluster.

Figure 13: Case 2: Nested Anti-Resonant (High Loss). All systems regress because a coordinated dual-gap adjustment is required.

Figure 14: Case 3: Hollow-Core ARF (High Loss). A single-ring topology reproduces the case 2 failure, ruling out nested geometry as the sole cause.

Figure 15: Case 4: Nested Anti-Resonant (Moderate Loss). Partial recovery (2/11 improve).

Figure 16: Case 5: Hybrid Plasmonic (High Loss). Occasional, unreliable success (2/11).

Figure 17: Case 6: Kagome (High Loss). Strongest success in the catalog: 9/10 systems reach the numerical floor through different edits.

Figure 18: Case 7: Multi-Core (High Loss). Coupling reasoning fails (1/11 marginal improvement).

Figure 19: Case 8: Suspended Cobweb (High Loss). Largest baseline loss, yet no system improves, and most err by reducing strut thickness in the wrong regime.

Figure 20: Case 9: Hollow-Core PBG. Near-optimal baseline, so all edits regress.

Figure 21: Case 10: All-Solid Bandgap. Simulator-floor baseline leaves no headroom.

Figure 22: Case 11: Solid-Core Hexagonal. Only one anomalous improvement.

Figure 23: Case 12: Graded-Index Multi-Ring. Part of the near-optimal cluster, with 0/11 improving.

Figure 24: Case 14: Vortex / Ring-Core. Near-floor baseline leaves no headroom (0/11 improve).

Figure 25: Case 15: High-Birefringence PM. Occasional success (1/11) but most systems regress.

Figure 26: Case 16: Elliptical Core. Most evenly contested case (6/11 improve).

Figure 27: Case 17: Square / Rectangular. One improvement, typical of the near-optimal cluster.

Figure 28: Case 18: Octagonal Lattice. Three improvements, the best at $-96.4$ dB.

Figure 29: Case 19: Twisted Hexagonal. Near-floor baseline, so every system regresses.

Figure 30: Case 20: Fractal / Quasi-Crystal. Without periodicity to exploit, every system regresses.

Inverse Design

Inverse design maps desired properties or behaviors to materials, geometries, molecules, and systems that can realize them.

See all 111 papers

May 28, 2026

Learning Design Skills as Memory Policies for Agentic Photonic Inverse Design

Photonic crystal fiber (PCF) inverse design remains challenging because candidate geometries must satisfy coupled optical targets under expensive electromagnetic simulation. Existing pipelines improve surrogate prediction or one-shot parameter recommendation, but they do not accumulate reusable design knowledge across iterative trials. We formulate PCF inverse design as a memory-policy learning problem and propose SkillPCF, a closed-loop agent framework that combines a physics-guided memory skill bank, reinforcement-learned skill selection, and simulator-grounded skill evolution. We further construct a real-world dataset with 479 expert interaction traces (2,507 spans) and 553 memory-dependent evaluation queries covering dispersion engineering, loss optimization, and multi-objective design. Experiments across multiple LLM backbones and classical baselines show that SkillPCF achieves stronger design-quality and efficiency trade-offs under practical simulation budgets, demonstrating the effectiveness of our proposed memory-skill learning paradigm for physics-aware PCF inverse design.

Shengchao Chen, Ting Shu, Sufen Ren

AAII, University of Technology Sydney · School of Artificial Intelligence, Shenzhen University · School of Information and Communication Engineering, Hainan University

LLM Agent Skill Learning Physics-Informed ML Memory-Augmented Agents Inverse Design RL for LLM Agents

Ask PDF

Aug 21, 2026

VortexChat: An agentic framework for autonomous multi-objective integrated photonic design

The advancement of modern integrated photonics is frequently bottlenecked by device design workflows that rely heavily on manual simulation and expert intuition. While inverse design offers an alternative, it remains constrained by expert supervision and a lack of end-to-end automation. To address these issues, we present VortexChat, an agentic framework for the autonomous, end-to-end inverse design of integrated photonic devices directly from natural language specifications. VortexChat couples a large language model (LLM) decision agent with topology generation, gradient-based refinement, and full-wave electromagnetic simulation. This closed-loop architecture enables the system to iteratively decompose design objectives, orchestrate computational tools, and update strategies based on feedback with minimal human intervention. Constrained by the absolute metrics of the Vortex100 Benchmark, VortexChat autonomously generates devices that strictly meet all predefined performance thresholds without any human-in-the-loop. As an experimental demonstration, we fabricated a broadband terahertz perfect vortex beam multiplexer, autonomously designed by VortexChat, with measurements confirming high-efficiency operation, high mode purity and low inter-channel crosstalk in agreement with full-wave simulations. These results demonstrate that an LLM agent can assume key aspects of expert decision-making in photonic inverse design while maintaining physical fidelity and fabrication feasibility, providing a scalable route towards autonomous design of complex integrated photonic systems.

Faqian Chong, Yulun Wu, Shilong Li +3

Innovative Institute of Electromagnetic Information and Electronic Integration, College of Information Science & Electronic Engineering, Zhejiang University, Hangzhou, 310027, China. · ZJU-Hangzhou Global Scientific and Technological Innovation Center, Zhejiang University, Hangzhou 310200, China. · International Joint Innovation Center, The Electromagnetics Academy at Zhejiang University, Zhejiang University, Haining 314400, China. +3

Inverse Design Tool-Augmented Language Model Agents LLM Agent Orchestration

Ask PDF

Aug 3, 2026

PICopilot: An LLM-based Agentic Framework for Assisting Photonic Integrated Circuit Design via Script Generation

The rapid development of photonic integrated circuits (PICs) is shifting the design flow from traditional graphical user interface (GUI)-based methods to script-based methods for higher flexibility, portability, and maintainability. However, script-based design introduces new challenges, requiring designers to possess additional proficiency in tool application programming interfaces (APIs) and programming. It also demands greater effort and time because it is inherently less intuitive and more complex than GUI-based methods. As PICs grow in scale and complexity, the productivity gap between design needs and manual scripting capabilities continues to widen. To address this gap, we introduce PICopilot, the first large language model (LLM)-based agentic framework that assists in PIC design via automated design script generation from natural language instructions. PICopilot leverages a multi-agent architecture with a feedback mechanism and a specifically designed retrieval-augmented generation (RAG) pipeline, achieving a high success rate and reliability. Experimental results on a benchmark of diverse PIC scripting tasks demonstrate that PICopilot successfully completes all 48 tasks and outperforms other LLM-based approaches without incurring substantial extra latency or cost, even solving 21 more tasks than the advanced GPT-5 model with a general RAG pipeline.

Xiaohan Jiang, Zeyu Li, Wei Zhang +1

Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology · 2Microelectronics Thrust, The Hong Kong University of Science and Technology (Guangzhou)

LLM-Based Program Synthesis

Ask PDF

Abstract

Inverse Design

Explore similar work

Learning Design Skills as Memory Policies for Agentic Photonic Inverse Design

VortexChat: An agentic framework for autonomous multi-objective integrated photonic design

PICopilot: An LLM-based Agentic Framework for Assisting Photonic Integrated Circuit Design via Script Generation