acceptodds
Under review as a conference paper at ICLR 2027

OptiGen: A Physics-Grounded Agentic Framework for Deep-Ultraviolet Laser Diode Design and Optimization

Abstract

The rapid emergence of agentic AI across scientific domains calls for extending its capabilities to physics-grounded design, where autonomous reasoning must be coupled with physical constraints and optimization. We presented OptiGen, an agentic framework for deep-ultraviolet laser diode (DUV LD) design and optimization. Unlike conventional LLM-based scientific assistants that primarily retrieve and summarize domain knowledge, OptiGen couples simulation-grounded designs with physics-grounded reasoning and constraint-aware optimization within a unified agentic loop. The framework performs query understanding, constraint-aware design retrieval, multi-objective optimization, physics-grounded reasoning, and grounded response generation. We evaluate OptiGen on benchmark queries covering single-objective, multi-objective, constraint-based, and worst-case design tasks using four models, namely o3, GPT-5, GPT-4, and Qwen, to assess the robustness of the framework. The OptiGen (o3) achieves a Precision@5 of 0.588, Recall@5 of 0.600, Mean Reciprocal Rank (MRR) of 0.640, and Normalized Discounted Cumulative Gain (NDCG)@5 of 0.612 for retrieval; best-design accuracy of 0.580 and performance accuracy of 0.659 for optimization; and a physics reasoning score of 0.830, response completeness of 0.884, and hallucination-free rate of 0.709 for generation. It further achieves an intent understanding accuracy of 0.890 and an end-to-end success rate of 0.570 in an average response time of 26.39 s. Across the evaluated models, OptiGen (o3) achieves the strongest overall performance across the major retrieval, optimization, and generation metrics, while OptiGen (Qwen) demonstrates competitive performance on intent understanding and retriever metrics.These results highlight the robustness of OptiGen across different language models demonstrating its potential as an autonomous decision system for specialized optoelectronic engineering problems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Agentic Systems, Optoelectronic Device Optimization, Physics-Informed Machine Learning, Deep-Ultraviolet Laser Diodes, Retrieval-Augmented Generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Q2hPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents OptiGen, an agentic framework for deep-ultraviolet laser diode design that combines simulation-based retrieval, physics knowledge retrieval, multi-objective optimization, and LLM-based reasoning. The framework is evaluated on 100 benchmark queries using four language models. The results demonstrate its potential for assisting device design, although further validation is needed to establish its effectiveness and the advantages of its agentic architecture.

Critical strengths
  1. The paper addresses a practical problem in DUV-LD design by integrating simulation-based retrieval, physics knowledge retrieval, multi-objective ranking, and LLM-based reasoning into a unified framework.
  2. The evaluation covers multiple design tasks and four language models, providing an initial assessment of the framework's performance across different settings.
Critical weaknesses
  1. The framework mainly follows a fixed workflow. The paper does not clearly explain how the agents work together or why this design is better than a simpler pipeline.
  2. Removing the simulation retrieval module improves several key metrics, including the hallucination-free rate and end-to-end success rate. This raises questions about whether the module is needed.
  3. The benchmark is limited, and the paper does not clearly explain how the ground truth is built. The expert evaluation includes only five queries, which is not enough to show that the system is reliable in real design tasks.
  4. The paper tests only four language models, some of which are relatively old. Testing newer models could help show whether the framework remains useful with stronger models.
Questions for authors
  1. The framework seems to follow a fixed workflow. Can the authors explain what the agentic design adds compared with a simple pipeline?
  2. Why does the performance improve when the simulation retriever is removed? It would be helpful if the authors could explain why this module is needed.
  3. How are the ground-truth results for the test queries obtained? Could the authors also test more recent and advanced models, such as Claude Opus, and include more expert evaluations to better validate the system?
1 comment

We sincerely thank the reviewer for the constructive feedback and for recognizing the practical relevance of OptiGen to DUV laser diode design. We have carefully considered the concerns and made additional clarifications and experimental revisions to strengthen the methodological explanation and evaluation of the framework.

1. Agentic architecture and advantages over a conventional pipeline

OptiGen is designed as a domain-specific, LLM-orchestrated framework for AlGaN deep-ultraviolet laser diode (DUV-LD) design. Its objective is to improve the precision and scientific reliability of design recommendations by grounding LLM reasoning in simulation data and domain-specific scientific knowledge.

The framework integrates two complementary knowledge sources: a SiLENSe-generated simulation database containing DUV-LD structures targeting 267 nm emission, and a physics knowledge base covering the materials, structural properties, and operating mechanisms of AlGaN-based DUV-LDs. Simulation retrieval identifies relevant device configurations and performance characteristics, while physics retrieval provides the scientific principles needed to interpret them. The parsed query determines the execution of task-specific components, including design ranking for optimization queries. Supplementary web retrieval is conditionally activated when simulation results do not satisfy the specified requirements or the available physics context is insufficient.

We clarify that OptiGen is not intended to provide unrestricted autonomous planning. Rather, it coordinates specialized components through a domain-specific workflow. Its contribution lies in integrating simulation-grounded evidence, physics-informed retrieval, task-dependent ranking, and conditional knowledge acquisition. We have also expanded the comparative evaluation by reporting expert and LLM-as-a-judge assessments against standalone general-purpose models, including ChatGPT and Claude.

2. Simulation Retriever ablation results

We acknowledge the reviewer’s observation that removing the Simulation Retriever improves certain metrics, particularly the hallucination-free rate and end-to-end success rate. This suggests that the current implementation may introduce redundant or less relevant simulation-derived context, especially because the downstream Ranker independently accesses the same simulation dataset.

Although both components use the same dataset, they serve distinct purposes. The Simulation Retriever identifies relevant simulated structures and their performance characteristics to support evidence-based reasoning and explanations. In contrast, the Ranker selects candidates through Pareto filtering and weighted ranking based on optical output power, threshold current, and optical confinement factor.

The observed improvement after removing the retriever may therefore result from overlapping information that does not always benefit response generation. Nevertheless, the retriever provides supporting simulation evidence, particularly for queries requiring the identification and interpretation of relevant structures. We have clarified these distinct roles and added a discussion of the observed ablation results and potential information redundancy to the limitations section.

3. Ground-truth construction, advanced LLM evaluation, and expert assessment

The benchmark ground truth was established through manual verification using the SiLENSe simulation dataset as the primary reference, together with domain knowledge of AlGaN-based DUV-LDs. We examined the simulated structures and their corresponding optical output power, threshold current, and optical confinement factor. Reference structures and rankings were verified according to each query’s optimization objectives and constraints, ensuring that expected results were grounded in simulation evidence and relevant physical principles rather than solely in LLM-generated responses.

To evaluate a more recent reasoning model, we incorporated Claude Opus 5.5 by replacing the underlying LLM within the existing OptiGen framework while keeping the architecture, retrieval pipeline, benchmark queries, and evaluation protocol unchanged. This provides a controlled comparison of different reasoning backbones within the same framework.

Additionally, we conducted a separate comparative evaluation of OptiGen against standalone general-purpose models including ChatGPT and Claude, on 10 queries. We have incorporated the corresponding results in the revised manuscript.

Reviewer_gUkpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OptiGen, in which an LLM parses a natural-language query into a structured intent, and a ranker selects top candidates from a database of 185 SiLENSe-simulated structures via Pareto filtering and weighted scoring. A physics retriever over DUV-LD literature and simulator documentation supplies supporting evidence, with web retrieval as a fallback, and a reasoning model produces the final design recommendation with physics-based explanations. Experiments with four LLM backbones on a 100-query benchmark demonstrate the effectiveness of the proposed method.

Critical strengths
  • This paper considers an interesting research question about LLM-based DUV laser diode design.

  • The author provides details implementation details and prompts in the appendix.

Critical weaknesses
  • The Acknowledgments name the authors' institution, violating double-blind anonymity. So this paper should not be further reviewed.

  • Selection reduces to a hand-fixed score that ignores the objectives parsed from the query and omits the emission wavelength.

  • The multi-agent framework is a fixed linear pipeline (one LLM call to parse, deterministic retrieval and ranking, one call to write) with no planning, tool choice, or iteration, and the parser's routing flags are never used because Alg. 1 always calls both retrievers. It would be better to check these carefully.

Questions for authors

Please see my Weaknesses.

1 comment

We sincerely thank the reviewer for the careful assessment and constructive feedback. We have addressed the concerns through manuscript revisions, additional experiments, and clarifications of the framework's execution logic.

Regarding double-blind anonymity: We apologize for this oversight. The acknowledgments section was unintentionally included in the submitted manuscript. We have removed this section from the revised manuscript to ensure compliance with the double-blind review requirements.

Regarding query-adaptive objective selection and emission wavelength: We thank the reviewer for this valuable suggestion. In response, we implemented and evaluated a query-adaptive objective selection and ranking strategy that dynamically considers the optimization objectives specified in each natural-language query. We compared this approach against the original fixed-weight ranking strategy. Although the query-adaptive approach provides objective-specific ranking, our experiments showed that the original fixed-weight strategy achieved slightly better overall performance across the benchmark. Therefore, we retained the fixed-weight strategy in the main framework and included the additional experimental results for query-adaptive ranking in the appendix. We have also clarified the rationale for this design choice in the revised manuscript. Regarding emission wavelength, our study specifically focuses on AlGaN-based deep-ultraviolet laser diodes operating at a target emission wavelength of 267 nm. Consequently, the emission wavelength is treated as a fixed design constraint rather than an optimization objective, and we have explicitly clarified this scope in the revised manuscript.

Regarding the framework orchestration, conditional execution, and planning capabilities: We appreciate the reviewer's observation and agree that the original description did not sufficiently clarify the execution logic of the framework. Both simulation-data retrieval and physics-knowledge retrieval are intentionally performed for every query because they provide complementary evidence. Simulation retrieval identifies relevant DUV-LD configurations and their simulated performance, whereas physics retrieval provides the underlying physical principles and mechanisms needed to interpret the retrieved structures and support scientifically grounded recommendations. Since physics-grounded reasoning is a central objective of OptiGen, physics retrieval is treated as a mandatory component rather than an optional operation.

However, the parsed task determines the subsequent execution of the framework. Specifically, the design-ranking module is activated for optimization-related queries, whereas it is not required for queries that only request information retrieval or explanation. Furthermore, supplementary web-based knowledge retrieval is conditionally triggered when the simulation results indicate that no retrieved design satisfies the specified requirements or when the available physics context is insufficient to support the requested explanation. The retrieved evidence is then integrated into a unified prompt for the final response generation.

We also acknowledge the reviewer's distinction between a coordinated agentic workflow and a fully autonomous agent with iterative planning. The current OptiGen implementation follows a structured orchestration pipeline comprising query interpretation, specialized retrieval components, conditional module execution, design ranking, and LLM-based response generation. It does not implement iterative autonomous planning, feedback-driven replanning, or unrestricted dynamic tool selection. We have revised the methodological description and Algorithm 1 to explicitly distinguish mandatory retrieval operations from conditionally executed modules, clarify how the parsed task influences downstream execution, and accurately describe the framework's current orchestration capabilities.

Reviewer_8QjYOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed an agentic framework for deep-ultraviolet laser diode design and optimization. OptiGen includes simulation grounded designs that consider physics reasoning and contact-aware optimization. Instead of generating new designs, OptiGen retrieves and ranks existing simulated designs with physics-grounded explanations for its output.

Critical strengths

None

Critical weaknesses
  • W1: No method-wise novelty in agentic framework specifically for deep-ultraviolet laser diode design. While the task coverage is comprehensive but does not highlight the framework design itself. The paper seems more suitable for physics or chemistry journal.

  • W2: The self-generated benchmark query has limited scope. The LLM-as-a-judge evaluation of self-generated queries is poorly documented. The authors do not clarify whether scientific correctness is validated against ground-truth evidence from physics literature or provide a clear scoring rubric that distinguishes, for example, a score of 3 from 4. The baselines are comparing with variants of OptiGen. Would love to see the comparison with recent frontier LLM agents with minimal additional scaffolding, such as Astra or Opus 5.5 with web-search capabilities, to better understand the contribution of the proposed framework.

Questions for authors
  • What’s the source of benchmark query? What sources were used, and was there any human involvement in generating or validating these queries?

  • How were the recent physics publications selected for the knowledge base? Also, how much human effort was required to curate and maintain this knowledge base?

1 comment

We sincerely thank the reviewer for the careful assessment and constructive feedback.

Benchmark query construction and validation: The benchmark queries were developed based on literature-informed design and optimization challenges in DUV laser diodes (LDs). We first examined relevant DUV-LD literature to identify the fundamental design objectives, performance trade-offs, and structural considerations commonly addressed by researchers and device designers. Based on these engineering considerations and insights from the literature, we manually defined 10 query templates, each comprising multiple natural-language formulations. These templates cover single-objective optimization (optical output power, threshold current, and optical confinement factor (OCF)), worst-case design identification, Pareto-based optimization, constraint-based retrieval, and joint optimization involving multiple performance metrics. The scientific context and evaluation criteria were further grounded in our SiLENSe simulation dataset. Human involvement primarily occurred during the identification of literature-informed design challenges and the formulation of the corresponding query templates. Thus, the benchmark represents domain-informed design and optimization scenarios. We have clarified the query-generation procedure and the associated human involvement in the revised manuscript and supplementary material.

Physics knowledge-base construction and human involvement: The physics knowledge base was curated through a domain-specific literature selection process, focusing on publications directly relevant to the physical mechanisms and structural design of AlGaN-based DUV LDs. Specifically, we selected publications addressing device performance improvements, structural optimization strategies, and approaches to enhance optical output power and optical confinement while reducing threshold current. The knowledge base was further enriched with the SiLENSe software manual, physics summaries, software release notes, and technical tutorials to incorporate device-simulation procedures, underlying physical principles, and practical guidelines for DUV-LD design and optimization.

Following the selection of relevant resources, knowledge-base construction was largely automated. PDF documents were processed using PyMuPDF to extract textual content and figures at the page level. The extracted figures were further analysed using o3 Vision to capture relevant physical information that might not be adequately represented in the extracted text. The resulting page-level representations were embedded using the BAAI/bge-large-en-v1.5 embedding model and indexed in a persistent ChromaDB collection, together with their associated publication and page metadata.

Human effort was primarily concentrated on selecting relevant scientific literature and technical resources to ensure adequate coverage of the physical mechanisms and design considerations governing DUV-LD performance. Additionally, the extracted information from the publications, SiLENSe manual, physics summaries, release notes, and tutorials was manually reviewed and verified by the authors for scientific relevance and consistency. We have included the selected publication list, together with the software manual and other technical resources, in the appendix for reference. We have also clarified the extent of human involvement and explicitly identified the knowledge-base construction tasks that were automated in the appendix.

Additional LLM comparisons: We sincerely thank the reviewer for suggesting comparisons with recent frontier LLMs and agentic systems. In response, we extended our evaluation by incorporating Claude Opus 5.5 into our experiments. First, we replaced the underlying reasoning LLM within the existing OptiGen framework with Claude Opus 5.5 while keeping the system architecture, retrieval pipeline, benchmark queries, and evaluation protocol unchanged. This controlled comparison enables us to examine the effect of the underlying reasoning model under identical framework conditions. The corresponding results are reported in Table 2.

Additionally, we conducted a separate comparative evaluation of OptiGen against general-purpose LLMs, including ChatGPT and Claude, as standalone baselines, as reported in Table 5. These complementary experiments provide insights into both the influence of the underlying reasoning LLM within OptiGen and the performance of the proposed domain-specific agentic framework relative to general-purpose language models. Together, these additional evaluations strengthen the empirical assessment of OptiGen and provide a clearer understanding of its contribution beyond the capabilities of the underlying LLM.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.