CoSP: Reconfigurable Multi-State Metamaterial Inverse Design via Contrastive Pretrained Large Language Model

TL;DR

CoSP tackles the challenge of inverse designing reconfigurable multi-state metamaterials by integrating a contrastive pretraining regime over multi-state spectra with a spectrum-guided large language model for material generation. The method first encodes spectra from multiple states into a unified latent space using a Transformer-based spectrum encoder trained with InfoNCE loss, then describes target spectra in natural language and completes design with a GPT-2–initialized Transformer decoder through cross-attention. Key contributions include (i) a two-stage contrastive-spectrum+LLM framework, (ii) a spectrum-description–driven inverse design to improve generalization to unseen spectra, and (iii) demonstration on VO$_2$-based RMMs, including hypothetical square-wave spectra for IR camouflage, with results indicating accurate spectrum matching and robust generalization beyond training data. The approach provides a Maxwell-equation–aware, autoregressive design flow that can explore non-existent spectra, enabling versatile and adaptive metamaterial design with practical impact in photonics and thermal management.

Abstract

Metamaterials, known for their ability to manipulate light at subwavelength scales, face significant design challenges due to their complex and sophisticated structures. Consequently, deep learning has emerged as a powerful tool to streamline their design process. Reconfigurable multi-state metamaterials (RMMs) with adjustable parameters can switch their optical characteristics between different states upon external stimulation, leading to numerous applications. However, existing deep learning-based inverse design methods fall short in considering reconfigurability with multi-state switching. To address this challenge, we propose CoSP, an intelligent inverse design method based on contrastive pretrained large language model (LLM). By performing contrastive pretraining on multi-state spectrum, a well-trained spectrum encoder capable of understanding the spectrum is obtained, and it subsequently interacts with a pretrained LLM. This approach allows the model to preserve its linguistic capabilities while also comprehending Maxwell's Equations, enabling it to describe material structures with target optical properties in natural language. Our experiments demonstrate that CoSP can design corresponding thin-film metamaterial structures for arbitrary multi-state, multi-band optical responses, showing great potentials in the intelligent design of RMMs for versatile applications.

Figures (6)

• Figure 1: Schematic diagrams illustrating (a) the concept of inverse design, where metamaterial structures are designed based on desired physical responses, (b) the concept of RMMs, which exhibit distinct changes in physical properties (thermal emissivity in this example) before and after state switching, and (c) the limitation of existing inverse design methods that can only aim at one single state, rather than accommodating multiple states.
• Figure 2: Architecture of the proposed CoSP framework for generating RMMs with multi-state optical responses in free text sequence.
• Figure 3: CoSP inverse design of RMMs for general purposes. Target Spectrum Pair represents the input spectrum with two states. Generated Text comprises the natural language results produced by CoSP, including the spectrum description and material structure sequence of the designed RMM. The spectrum description, highlighted in color, corresponds to the wavelength bands depicted in the first column's spectra. Spectrum of Designed RMMs illustrates the spectrum calculated using TMM for the material structure generated by CoSP. Additionally, the sample in the training set with the closest match to the target spectrum is displayed, along with the MSE calculated for both, to demonstrate the generalization capability of CoSP.
• Figure 4: Illustration of inverse design results aimed at IR-camouflage.
• Figure 5: Visualization of raw spectrum (left) and spectrum representation (right) after UMAP dimensionality reduction. Three pairs of representative spectrum pairs are highlighted as point pairs with different colors (red, purple, and blue).