奈米光子結構的反向設計任務具有高度非線性與一對多解空間的特性，傳統優化方法在多樣性與效率上往往受限。因此，本研究引入生成式深度學習模型以實現同時具備結構多樣性與光學性能要求的結構生成。本文聚焦於 Taigao Ma 等人所提出之論文的 Task 2，針對自由形狀矽結構在 400–680 nm 可見光範圍內之穿透光譜進行反向設計。為進一步探討生成模型於實際應用需求中的效能，本研究選擇條件變異自編碼器（Conditional Variational Autoencoder, cVAE）與條件生成對抗網路（Conditional Generative Adversarial Network, cGAN）兩種典型模型進行比較。
模型訓練完成後，分別進行三種生成任務：（1）直接以原始訓練資料進行生成；（2）篩選出 gap 值落在 250–350 nm 範圍內之結構；（3）篩選出在 520–560 nm 波段內具有穿透率高於 0.45 的光譜峰值之結構。共計產出六組實驗數據，加上forward model 的聯合訓練，有效提升模型對於物理可實現性的掌握度。
最後針對模型的光譜準確性、多樣性與條件滿足率進行評估與分析。實驗結果顯示，VAE 模型在無條件生成任務中表現穩定，具備良好的光譜準確度與運算效率，但在限制條件下較易出現 decoder 模式收斂的現象，導致生成結構趨於單一，降低了多樣性。相較之下，GAN 模型即使在嚴格條件下仍保有高結構多樣性與良好光譜匹配能力，適合應用於需滿足特定設計條件的場景。
本研究成果可為未來實務應用中的條件式光學設計提供具體依據，並作為不同生成模型選擇上的參考指標。

This paper addresses the challenges of inverse design in nanophotonic structures, where high structural degrees of freedom lead to high computational costs and low search efficiency. Traditional methods rely heavily on parameter sweeps and optical simulations, which become increasingly inefficient when dealing with free-form structures. To overcome these limitations, we adopt generative deep learning models to perform conditional inverse design. These models can directly generate structural images and periodic parameters (gap) from a target transmission spectrum. A pretrained forward neural network is also employed to verify the optical consistency of the generated results.
We develop two types of generative models: the Variational Autoencoder (VAE) and the Generative Adversarial Network (GAN), and evaluate them under three constraint settings: unconditional generation, gap-constrained generation, and spectrum-constrained generation. Model performance is systematically analyzed using metrics such as SSIM, RMSE, and structural irregularity.
Experimental results show that the VAE performs stably in the unconditional task, offering high spectral accuracy and fast inference. However, under stricter constraints, the decoder tends to collapse to a limited set of patterns, resulting in reduced diversity of generated structures. In contrast, the GAN model maintains high structural diversity even under strong constraints and achieves better spectral alignment, making it more suitable for applications that require meeting specific optical design criteria.
In summary, this study demonstrates the effectiveness of deep generative models for the inverse design of free-form nanophotonic structures. Depending on the task requirements, one can choose between VAE and GAN to balance generation efficiency and structural flexibility.

[1] S. Molesky, Z. Lin, A. Y. Piggott, W. Jin, J. Vučković, and A. W. Rodriguez, “Inverse design in nanophotonics,” Nature Photonics, vol. 12, no. 11, pp. 659–670, 2018.
[2] D. Liu, Y. Tan, E. Khoram, and Z. Yu, “Training deep neural networks for the inverse design of nanophotonic structures,” ACS Photonics, vol. 5, no. 4, pp. 1365–1369, 2018.
[3] X. Han, Z. Y. Fan, Z. Y. Liu, C. Li, and L. J. Guo, “Inverse design of metasurface optical filters using deep neural network with high degrees of freedom,” InfoMat, vol. 3, no. 4, pp. 432–442, 2021.
[4] T. Ma, Z. Liu, Y. Liu, and A. Majumdar, “Benchmarking deep learning-based models on nanophotonic inverse design problems,” Opto-Electronic Science, vol. 1, p. 210012, 2022.
[5] J. Jiang, D. Sell, S. Hoyer, J. Hickey, J. Yang, and J. A. Fan, “Free-form diffractive metagrating design based on generative adversarial networks,” ACS Nano, vol. 13, no. 8, pp. 8872–8878, 2019.
[6] Z. Liu, D. Zhu, S. P. Rodrigues, K.-T. Lee, and W. Cai, “Generative model for the inverse design of metasurfaces,” Nano Letters, vol. 18, no. 10, pp. 6570–6576, 2018.
[7] M. G. Moharam and T. K. Gaylord, “Rigorous coupled-wave analysis of planar-grating diffraction,” JOSA, vol. 71, no. 7, pp. 811–818, 1981.
[8] M. Born and E. Wolf, Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light, 7th ed. Cambridge University Press, 1999.
[9] S. A. Maier, Plasmonics: Fundamentals and Applications. Springer, 2007.
[10] “Spectral sensitivity,” Wikipedia. [Online]. Available: https://en.wikipedia.org/wiki/Spectral_sensitivity
[11] K. Suzuki, “Artificial neural networks - methodological advances and biomedical applications,” IntechOpen, pp. 3–68, 2011.
[12] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, pp. 533–536, 1986.
[13] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proc. Int. Conf. Learn. Representations (ICLR), 2015.
[14] Z. Li, F. Liu, W. Yang, and S. Peng, “A survey of convolutional neural networks: analysis, applications, and prospects,” Unpublished.
[15] D. P. Kingma and M. Welling, “Auto-encoding variational Bayes,” in Proc. Int. Conf. Learn. Representations (ICLR), 2014.
[16] M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv preprint arXiv:1411.1784, 2014.
[17] T. Ma, Z. Liu, Y. Liu, and A. Majumdar, “Supplementary information for ‘Benchmarking deep learning-based models on nanophotonic inverse design problems’,” Opto-Electronic Science, vol. 1, p. 210012, 2022.
[18] A. Malkiel et al., “Plasmonic nanostructure design and characterization via deep learning,” Light: Science & Applications, vol. 7, 2018.
[19] J. Jiang et al., “Deep neural networks for the evaluation and design of photonic devices,” Nature Reviews Materials, vol. 7, pp. 679–695, 2022.
[20] R. Yamashita, M. Nishio, R. K. G. Do, and K. Togashi, “Convolutional neural networks: an overview and application in radiology,” Insights into Imaging, vol. 9, pp. 611–629, 2018.
[21] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” arXiv preprint arXiv:1511.06434, 2015.