Liquid crystal microlens arrays (LC-MLAs) are widely used in spectral reconstruction, optical field imaging, and 3D reconstruction. However, current LC-MLA optimization approaches predominantly focus on improving the structural design and material properties to enhance optical performance. These methods typically follow a sequential paradigm when selecting the structural parameters of LC-MLAs, making it difficult to reach optimal solutions due to the discrepancies between designed and actual values. In addition, the lengthy iterative process and inefficiencies often result in low transmittance and insufficient beam-focusing abilities, compromising the quality of spectral reconstruction. To address these limitations, we propose a deep learning-based liquid crystal device inverse design network (LCDE-IDN) to optimize LC-MLA for hyperspectral reconstruction systems. Experimental results indicate that the LCDE-IDN method significantly enhances the transmittance and beam-focusing ability of LC-MLAs, thus improving the accuracy of spectral reconstruction. This provides an efficient and effective way to select optimal parameters for LC-MLA design.

In this paper, we propose the LCDE-IDN method, which integrates deep learning techniques with the physical characteristics of LC-MLAs to optimize structural parameters for hyperspectral reconstruction. The LCDE-IDN framework leverages a fully connected neural network and incorporates a pairwise learning strategy that combines a forward design network with an inverse design network. This approach enables a more effective capture of global features and nonlinear relationships between device structural parameters and the resulting spectral curves. Unlike traditional methods, the inverse design network does not directly learn from the original dataset; instead, it derives its network parameters from the pre-trained forward design network. This reduces the risk of overfitting and enhances accuracy, while also avoiding issues related to the non-uniqueness of spectral curve mappings to device parameters. Ultimately, the LCDE-IDN method delivers more precise structural parameters for LC-MLA and higher-quality spectral reconstruction compared to empirical methods.

To evaluate the performance of the LCDE-IDN method, we benchmark it against a conventional LC-MLA optimized design using the same materials, such as circular-hole electrodes, nematic-phase liquid crystal molecules, and polyimide alignment layers. Experimental results show that, compared to the LC-MLA designed by empirical methods, the LCDE-IDN-designed LC-MLA exhibits a reduction in light intensity uniformity error to approximately 4.5%, while average transmittance improves by 3.1%. These findings demonstrate that the LC-MLA optimized through LCDE-IDN possesses higher transmittance and enhanced beam-focusing abilities, outperforming empirically designed LC-MLAs in terms of optical field imaging performance (Fig. 5). In hyperspectral reconstruction, the LCDE-IDN-optimized LC-MLA improves the peak signal-to-noise ratio (PSNR) of reconstructed spectral images by an average of 5.7% (Fig. 6). We also analyze the effect of LC-MLA fabrication errors on spectral reconstruction results (Table 1). The results indicate that the structural parameters designed by the LCDE-IDN method provide a buffer against fabrication inaccuracies, enabling a quantitatively optimized LC-MLA design.

We propose a deep learning-based optimization method for LC-MLA design in hyperspectral reconstruction systems. Through optoelectronic performance testing, hyperspectral reconstruction experiments, and discussions on fabrication errors, it is shown that the LCDE-IDN method can accurately achieve optimal LC-MLA structural designs. The optimized LC-MLA demonstrates superior spectral image reconstruction compared to empirical methods. As a result, the LCDE-IDN method overcomes the limitations of traditional LC-MLA design approaches, significantly improving both the optoelectronic performance and spectral reconstruction capabilities of LC-MLAs. This advanced LC-MLA technology shows promising applications in agriculture, medicine, and chemical industry.