PaperlayerView exact revision on arXiv

Source-linked AI summary

An Agentic Framework for Autonomous Metamaterial Modeling and Inverse Design

Darui Lu, Jordan M. Malof, Willie J. Padilla

arXiv:2506.06935v2cs.AIcond-mat.mtrl-sci

TL;DR

Human-led inverse design entails numerous decisions and can be time-consuming. The paper presents an agentic framework for automated metamaterial inverse design, and reports performance comparable to human expert-designed solutions. Its stated future work is to give the framework the ability to create new tools and agents itself.

  • Problem

    Human researchers make numerous design choices in forward and inverse modeling, and the process requires substantial time.

    Source: §1 · p. 2

  • Method

    The framework automates metamaterial inverse design, using a Planner and an autonomous agent in the end-to-end design process.

    Source: §4 · p. 18

  • Results

    The framework consistently achieves performance comparable to human expert-designed solutions.

    Source: §4 · p. 18

  • Takeaways & Limitations

    The authors identify the framework's automation capabilities and potential to accelerate deep learning as a supported conclusion.

    Source: §4 · pp. 18, 20

    Self-improvement through creating new tools and agents is left for future work.

    Source: §4 · p. 20

Abstract

from arXiv · show

Recent significant advances in integrating multiple Large Language Model (LLM) systems have enabled Agentic Frameworks capable of performing complex tasks autonomously, including novel scientific research. We develop and demonstrate such a framework specifically for the inverse design of photonic metamaterials. When queried with a desired optical spectrum, the Agent autonomously proposes and develops a forward deep learning model, accesses external tools via APIs for tasks like simulation and optimization, utilizes memory, and generates a final design via a deep inverse method. The framework's effectiveness is demonstrated in its ability to automate, reason, plan, and adapt. Notably, the Agentic Framework possesses internal reflection and decision flexibility, permitting highly varied and potentially novel outputs.

1 Introduction

Human-led metamaterial modeling and inverse design require substantial planning, tool, algorithm, and implementation choices, consuming time and expertise. The paper introduces an agentic framework to automate the end-to-end inverse-design process, including development and optimization of a forward model, with specialized agents, internal reflection, and dynamic planning.

1 Introduction

  • Human researchers must plan simulations, choose tools and algorithms, implement procedures, and iteratively adjust them, making the process time-consuming and expertise-intensive.The authors say these demands slow scientific progress and limit access to advanced pipelines.Source: §1 · p. 2
  • The framework automates end-to-end metamaterial inverse design, including autonomous development and optimization of a DNN-based surrogate forward model.Source: §1 · p. 3
  • Specialized LLM-based agents collaborate on complex scientific tasks, with performance described as similar to that of a human scientist.Source: §1 · p. 3
  • Internal reflection and dynamic planning let the framework adapt its methodology, supporting a more autonomous research process.Source: §1 · p. 3

2 The Agentic Framework

The Agentic Framework automates photonic metamaterial inverse design by coordinating task planning and input checks with forward-model training and inverse design. Its workflow uses a trained surrogate and can optionally verify the selected geometry through electromagnetic simulation.

2 The Agentic Framework

  • A dedicated Planner coordinates user communication, LLM management, and tool use, while specialized LLM modules and memory support task execution over time.The framework uses a built-in memory store to record and retrieve chat history; its modular design is intended to ease development, debugging, and extension.Source: §2 · p. 5
  • The framework supports the full process from forward-model training to inverse design and returns a metamaterial geometry and material description intended to achieve the user's specified goal.The authors state that it can extend beyond the demonstrated photonic system to other material and spectroscopic systems.Source: §2 · p. 7
  • The authors demonstrate the framework on an established photonic metamaterial benchmark and report internal reflection, versatility, and inverse-design success.They state that the system is not limited to the demonstrated material system.Source: §2 · p. 7

2.1 Planner

  • The Planner gathers the target spectrum and task context, then divides the overall goal into tasks and assigns responsibility to agents.Possible plans include developing a forward model alone or developing one and performing inverse design.Source: §2.1 · pp. 7–8
  • Rather than following a fixed script, the Planner evaluates task outcomes and chooses subsequent steps, enabling the framework to adapt its course and produce varied, potentially novel results.It stores an initial high-level strategy in memory and iterates through planning, execution, and analysis.Source: §2.1 · p. 8
  • The Planner gives AIDE a task description specifying input/output dimensions, evaluation metrics, and possible solutions to guide controllable code generation.AIDE is an LLM coding agent for machine-learning tasks.Source: §2.1 · p. 8

2.2 Input Verifier

  • The Input Verifier checks required files and code inputs, including training-data sources for forward modeling and a target-spectrum file for inverse design, and prompts the user if inputs are missing.Its file-check tool tests whether supplied paths exist.Source: §2.2 · p. 8

2.3 Forward Modeler

  • The Controller iteratively balances data acquisition against model-architecture optimization, choosing to test, generate a model, or stop based on performance history and the target metric.The loop grows the dataset and ends when the target or convergence is reached, or when a hard-coded data or round limit is hit; the Controller records its reasoning.Source: §2.3 · p. 9, Figure 2 · p. 10

2.4 Inverse Designer

  • Neural Adjoint reuses the trained forward surrogate without retraining, generates candidate geometries by gradient optimization, and returns the geometry with the lowest prediction error against the target spectrum.The inverse task maps a desired spectral response to a metamaterial geometry.Source: §2.4 · pp. 11–12
  • A human designer achieved an inverse result more than two orders of magnitude better than AIDE, so the framework uses Neural Adjoint for inverse design.AIDE is described as accurate for forward prediction but unsuitable for inverse design.Source: §2.4 · p. 11
  • If a computational electromagnetic simulation tool is available, the Inverse Designer can verify the selected geometry using a user-supplied CST file and export a .cst design file.Figure 3 depicts this optional numerical-simulation step after Neural Adjoint optimization.Source: §2.4 · p. 12, Figure 3 · p. 13

3 Results and Discussion

The framework is evaluated for all-dielectric metamaterial inverse design through adaptive forward-model development and optimization with a fixed large dataset, with inverse design following forward-model development in both settings. The section also reports higher inverse error than human-designed inverse models and discusses fixed Neural Adjoint hyperparameters as a possible contributor.

3.1 Target MSE

  • At about 3,900 samples, diminishing MSE improvements led the Agent to regenerate model code, while worse code-generation results could prompt another regeneration.The Agent’s internal reasoning attributed the switch to diminishing returns and saturation.Source: §3.1 · p. 15
  • In four experiments, the Agent reached the forward target validation MSE of 2 × 10^-3, using dataset sizes around 20,000, 24,000, 11,500, and 15,000, with differing neural network architectures.Source: §3.1 · p. 15
  • Inverse-designed geometries achieved an average re-simulated MSE of 1.7 × 10^-3 against target spectra, corresponding to an average validation error of about 0.2%.The test set contained 100 target spectra.Source: §3.1 · p. 16

3.2 Fixed Large Dataset

  • A four-layer Transformer encoder followed by a three-layer residual multilayer perceptron, with a skip connection from its input to final hidden layer, achieved unseen-test MSE 1.3 × 10^-3.Source: §3.2 · p. 17
  • For the fixed-dataset inverse-design step, the Inverse Designer applied the NA tool and evaluated results on a 100-sample test set.The passage points to Table 1 and Figure 6 for inverse-design results but does not provide their values here.Source: §3.2 · p. 18
  • The Agentic system produces higher inverse error than human-designed inverse models [8,19].The authors attribute this gap to Neural Adjoint being implemented as a pre-coded tool with fixed hyperparameters, such as the number of backpropagation steps.Source: §3.2 · p. 18
  • The authors believe that giving the agent control to fine-tune the Neural Adjoint loop could improve inverse performance.The current Neural Adjoint tool uses fixed hyperparameters, including the number of backpropagation steps.Source: §3.2 · p. 18

4 Conclusion

The proposed framework automates inverse design of metamaterial systems through coordinated agents, and experiments on all-dielectric metasurfaces show adaptation across stages and performance comparable to human expert-designed solutions.

4 Conclusion

  • The framework automates inverse design using a Planner, Input Verifier, Forward Modeler, and Inverse Designer.These roles orchestrate the process, check input integrity, develop surrogate models, and propose geometries that achieve the goal.Source: §4 · p. 18
  • The framework dynamically adapts its strategy at each stage and autonomously controls the entire design process.The experiments were conducted on all-dielectric metasurfaces.Source: §4 · p. 18
  • The framework consistently achieves performance comparable to human expert-designed solutions.The authors report this in experiments on all-dielectric metasurfaces.Source: §4 · p. 18
  • The authors highlight the framework's potential to accelerate deep learning research in metamaterials and related areas.Source: §4 · p. 20
  • Future work will enable the framework to create new tools and agents as a step toward an Agentic Researcher.Source: §4 · p. 20

Supporting Information Available

The supporting information details the system prompts for the framework's four LLM components.

Supporting Information Available

  • The supporting information details the system prompts for the Planner, Input Verifier, Forward Modeler, and Inverse Designer LLMs.Each rounded rectangle in Figure 1 represents an LLM with a system prompt designed for the forward and inverse design process.Source: Supporting Information Available · p. 20
Loading 2506.06935v2…