Recently, deep learning has been used to establish the nonlinear and nonintuitive mapping between physical structures and electromagnetic responses of meta-atoms for higher computational efficiency. However, to obtain sufficiently accurate predictions, the conventional deep-learning-based method consumes excessive time to collect the data set, thus hindering its wide application in this interdisciplinary field. We introduce a spectral transfer-learning-based metasurface design method to achieve excellent performance on a small data set with only 1000 samples in the target waveband by utilizing open-source data from another spectral range. We demonstrate three transfer strategies and experimentally quantify their performance, among which the “frozen-none” robustly improves the prediction accuracy by ∼26 % compared to direct learning. We propose to use a complex-valued deep neural network during the training process to further improve the spectral predicting precision by ∼30 % compared to its real-valued counterparts. We design several typical teraherz metadevices by employing a hybrid inverse model consolidating this trained target network and a global optimization algorithm. The simulated results successfully validate the capability of our approach. Our work provides a universal methodology for efficient and accurate metasurface design in arbitrary wavebands, which will pave the way toward the automated and mass production of metasurfaces.

As a successful case of combining deep learning with photonics, the research on optical machine learning has recently undergone rapid development. Among various optical classification frameworks, diffractive networks have been shown to have unique advantages in all-optical reasoning. As an important property of light, the orbital angular momentum (OAM) of light shows orthogonality and mode-infinity, which can enhance the ability of parallel classification in information processing. However, there have been few all-optical diffractive networks under the OAM mode encoding. Here, we report a strategy of OAM-encoded diffractive deep neural network (OAM-encoded D²NN) that encodes the spatial information of objects into the OAM spectrum of the diffracted light to perform all-optical object classification. We demonstrated three different OAM-encoded D²NNs to realize (1) single detector OAM-encoded D²NN for single task classification, (2) single detector OAM-encoded D²NN for multitask classification, and (3) multidetector OAM-encoded D²NN for repeatable multitask classification. We provide a feasible way to improve the performance of all-optical object classification and open up promising research directions for D²NN by proposing OAM-encoded D²NN.

从极少量的测量值中有效且高概率高质量恢复出原始信号是压缩感知图像重建研究的核心问题，学者们相继提出了传统和基于深度学习的压缩感知图像重建算法，传统算法通常基于优化模型迭代求解，重建质量和重建速度都无法保证；基于深度学习的算法重建质量相对较高，但缺乏物理可解释性。受滤波流的启发，本文提出了联合全局与局部的深度压缩感知图像重建模型（G2LNet），其以卷积层执行压缩采样以及初始重建过程，利用快速傅里叶卷积与滤波流，同时考虑了图像全局上下文信息和图像像素局部邻域信息，联合学习优化测量矩阵与滤波流，建立了完整的端到端可训练的深度图像压缩感知重建网络。经实验验证，在压缩感知图像重建领域常用的Set5，Set11，BSD68测试集上取得了良好的重建效果，在采样率为20%的情况下，G2LNet的图像重建质量相比于经典的传统算法MH与基于深度学习的算法CSNet的平均PSNR分别提高了2.29 dB，0.51 dB，有效提升了重建图像质量。

高能重离子碰撞将夸克和胶子从原子核中释放出来，形成一种新的核物质形态，即解禁闭的夸克胶子等离子体（Quark-Gluon Plasma，QGP）。研究普通核物质或强子共振气体到夸克胶子等离子体的相变，需要在超级计算机上数值求解格点量子色动力学（Quantum Chromodynamics，QCD）。但是，格点QCD只能给出零重子化学势以及附近可泰勒展开区域的核物质状态方程，并预测这种条件下QGP到强子共振气体之间的相变为平滑过渡。在不能做泰勒展开的有限重子化学势区域，格点QCD会遭遇著名的符号问题，无法给出有效的核物质状态方程以及QCD相变类型。本文综述了利用机器学习在核物质状态方程、相变分类以及临界点寻找方面的研究。这些研究大致分为两类：第一类在高能重离子碰撞实验数据以及相对论流体力学模拟和分子动力学模拟中，利用核碰撞末态粒子分布来确定核物质状态方程以及相变种类；另一类是利用机器学习直接帮助格点QCD的采样，解决有限密系统中的符号问题。