Skip to content
Open access

Conditional flow matching with composable dynamic constraint guidance for inverse design of terahertz metamaterials

Sep 2026 · Journal of Physics D: Applied Physics · Vol 59, pp. 385107 · 0 citations · 37 references
Physics

Abstract

Efficient inverse design remains challenging for terahertz metamaterials. Existing methods often struggle to balance generation quality and inference efficiency, and the functionality of trained models is relatively fixed, making it difficult to flexibly adapt to additional design requirements. To address these issues, this paper proposes a conditional flow matching (CFM) framework with composable dynamic constraint guidance (CDCG) for parametric inverse design of terahertz metamaterial structures. As the generative backbone, CFM learns a conditional velocity field under a given target response and enables the progressive evolution from a random noise distribution to the structural-parameter distribution, while maintaining generation quality and improving inference efficiency. CDCG is introduced to decouple basic spectral-indicator generation from additional-objective guidance. It does not require backbone retraining. During inference, lightweight constraint predictors provide gradients, guiding generation toward feasible regions that satisfy the added constraints. Compared with denoising diffusion probabilistic models, CFM achieves a 4.5-fold increase in inference speed while maintaining accurate mapping from spectral responses to structural parameters. Compared with schemes that require model retraining when new design objectives are introduced, CDCG reduces the adaptation time from 1540 s to 70 s while maintaining comparable performance on the primary design metrics. In addition, the proposed framework enables coordinated optimization under conflicting constraints involving electromagnetic performance, material consumption, and structural feasibility, providing a flexible strategy for multi-objective inverse design of terahertz metamaterial structures.

Read PDF

Similar papers

Open access Oct 2026

Autonomous Discovery of Multi-Material Metamaterials With Programmable Multimodal Thermal Responses via Reinforcement Learning.

Inverse design deep learning models have significantly accelerated materials discovery through mapping complex structure-property relationships. However, they remain constrained by prohibitive data demands and are limited to in-distribution interpolation designs. Efficient out-of-distribution inverse design under spars...

Chen-Xi Lu, Ke-Wen Pan, Xing-Ran Li et al. · 0 citations
Open access Nov 2026

Language-guided inverse design of multiclass metamaterials using generative artificial intelligence

MetaGen-LLM is introduced as a language-guided framework that maps free-form design requests to high-fidelity geometries across eight metamaterial classes, providing a unified route from natural language to continuous mesh representations of multiclass metamaterials.

Xiao-Yang Zheng · 0 citations
Conference Open access Oct 2026

Pose-driven neural radiance field fusion with geometric conformance for virtual try-on

Virtual try-on requires geometrically faithful garment reconstruction with pose-consistent dynamic fitting, but current methods struggle to achieve satisfactory reconstruction accuracy, controllability and computational efficiency. Mesh-based approaches miss out high-frequency garment geometry details, single implicit...

Yi-Jin Tang, Yu-Huan Lv, Zi-Sen Zhou et al. · 0 citations
Preprint Oct 2026

Flow Matching Reinforcement for 3D Mesh Generation via Dynamic Homing Optimization

Flow matching is central to 3D generation, yet in practice its reinforcement learning (RL) methods are largely adapted from 2D visual generation. Representative DPO-, GRPO-, and NFT-style objectives, when applied to negative trajectories, mainly steer predicted velocities away from the corresponding directions without...

Zhen-Wu-Yong Zhou, Zhi-Wei Ning, Pu-Hua Jiang et al. · 0 citations
Preprint Oct 2026

SLIM: Learning an Operator for Material-Efficient 3D Lattice Refinement

Lattice inverse-design methods can generate cells that match prescribed effective properties, but property feasibility does not imply material efficiency. Removing redundant material from an existing lattice while preserving its homogenized elasticity ordinarily requires solving a separate constrained topology optimiza...

Yue-Ze Zhu, Yu Xing, Tian-Yang Xue et al. · 0 citations
Preprint Oct 2026

Channel Match: Calibrating Shape Latents for Few-Step 3D Generation

This work introduces a few-step distillation framework for native 3D generation, enabled by continuous self-consistency training with a custom JVP FlashAttention kernel and an analytic velocity regression objective for sparse structured 3D latents.

Di Yang, Zhi-Hao Li, Yan-Hai Xiong et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.