Learning the Right Shapes: Dataset's Influence on AI-Assisted Computer-Generated Holography

This study examines how the quality and diversity of training data affect the performance of computer-generated holography assisted by artificial intelligence for laser beam shaping. Using the DeepCGH framework based on a convolutional neural network, we evaluate various datasets ranging from simple shapes to complex word-based patterns. The results show that diverse and heterogeneous data improve reconstruction accuracy and generalization. The work highlights the need for hybrid approaches that combine data diversity with physical modeling to achieve more stable and reliable holographic generation.