DNNsolver：偏振敏感衍射神经网络的精确高效深度学习建模
DNNsolver: Accurate and Efficient Deep Learning Modeling of Polarization-Sensitive Diffractive Neural Networks
- State Key Laboratory of Precision Measurement Technology and Instruments, Department of Precision Instrument, Tsinghua University（清华大学精密仪器系精密测量技术与仪器国家重点实验室）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
本文提出DNNsolver，一种基于深度学习的电磁替代模型，通过预测散射矩阵实现偏振敏感衍射神经网络的高效精确建模，显著加速模拟并优化多层分类器。
AI中文摘要：
衍射神经网络（DNNs）由级联的衍射层组成，为高速光学计算提供了一个有前景的平台。然而，对高密度、偏振敏感的DNNs进行精确且高效的建模仍然具有挑战性。基于薄元件近似（TEA）的传统模型无法捕捉层内的电磁耦合，而诸如时域有限差分（FDTD）方法等全波模拟计算成本高昂。在此，我们提出了DNNsolver，一种基于深度学习的电磁替代模型，用于学习衍射层的散射响应。通过预测散射矩阵而非依赖于源的输出场，DNNsolver将入射波前与模型解耦，并可应用于任意输入场。局部散射核和滑动窗口策略进一步允许对任意空间尺寸的DNNs进行高效建模。对于随机散斑照明下的单衍射层，DNNsolver实现了平均均方误差（MSE）为0.0095，并且与FDTD模拟相比获得了约10^6倍的加速。DNNsolver进一步应用于三层偏振复用图像分类器的优化，其与FDTD模拟的一致性优于基于TEA的对应方法。我们的工作为高密度、偏振敏感DNNs的设计和优化提供了一个精确且高效的框架。
英文摘要：
Diffractive neural networks (DNNs), composed of cascaded diffractive layers, offer a promising platform for high-speed optical computing. However, accurate and efficient modeling of high-density, polarization-sensitive DNNs remains challenging. Conventional model based on thin element approximation (TEA) fails to capture electromagnetic coupling within layers, while full-wave simulations such as finite-difference time-domain (FDTD) methods are computationally expensive. Here, we propose DNNsolver, a deep learning-based electromagnetic surrogate model that learns the scattering response of diffractive layers. By predicting the scattering matrix rather than source-dependent output fields, DNNsolver decouples the incident wavefront from the model and could be applied to arbitrary input fields. A local scattering kernel and sliding-window strategy further allow efficient modeling of DNNs with arbitrary spatial sizes. For single diffractive layer under random speckle illumination, DNNsolver achieves an average mean squared error (MSE) of 0.0095 and about 10^6-fold speedup compared with FDTD simulations. DNNsolver is further applied to the optimization of a trilayer polarization-multiplexed image classifier, which exhibits better agreement with FDTD simulations than its TEA-based counterpart. Our work provides an accurate and efficient framework for the design and optimization of high-density, polarization-sensitive DNNs.
