Self-referential holographic data storage (SR-HDS), which has been proposed as a novel implementation method for holographic data storage (HDS), enables holographic recording without a reference beam. In addition to the signal pattern (SP) to be recorded, an additional pattern (AP) that affects the reconstruction quality is used in SR-HDS. One of the methods for obtaining a designed AP that contributes to high-quality reconstruction involves utilizing local search algorithms, such as the hill climbing (HC) method. However, designing an AP using this method typically requires a significant amount of time. In this study, we proposed a new AP-designing method that uses a deep neural network. By training a network with pairs of SP and designed AP based on a local search algorithm, a designed AP that improves the reconstruction quality of an arbitrary SP can be instantly obtained. APs designed using the deep learning-based method improved the reconstruction quality of SPs to the same level as those designed using the method based on local search algorithm, whereas the time required to obtain one designed AP was reduced by three or four orders of magnitude.

自参考全息数据存储（SR-HDS）是一种新颖的全息数据存储（HDS）实现方法，可以在没有参考光束的情况下进行全息记录。除了要记录的信号模式（SP）之外，SR-HDS 中还使用影响重建质量的附加模式（AP）。获得有助于高质量重建的设计 AP 的方法之一涉及利用局部搜索算法，例如爬山 (HC) 方法。然而，使用这种方法设计 AP 通常需要大量时间。在这项研究中，我们提出了一种使用深度神经网络的新 AP 设计方法。通过基于局部搜索算法训练具有成对的 SP 和设计的 AP 的网络，可以立即获得提高任意 SP 的重建质量的设计的 AP。使用基于深度学习的方法设计的AP将SP的重建质量提高到使用基于局部搜索算法的方法设计的AP的相同水平，而获得一个设计的AP所需的时间减少了三到四个数量级。