Building upon controllable light-matter interactions, photonic metamaterials have demonstrated promising performances in sustainable applications spanning from UV to mid-infrared frequency ranges. These materials are complex, three-dimensional architected nanostructures with representative applications in state-of-the-art technologies such as solar panels, radiative cooling devices, and photocatalytic systems. However, efficient design and rapid verification of 3D photonic metamaterials remain an urgent challenge, significantly limiting operational efficiency and broader applicability in these domains. Hereby we aim at proposing a deep learning-based approach to enable highly-accurate, real-time, and wide-ranging design of 3D photonic metamaterials. The 3D structures will be generated via a metasuface-based holographic lithography technique, which leverages the spatial interference of multiple diffraction orders. We will begin by developing forward models to predict physical properties—such as the photonic density of states or complete photonic bandgaps—based on the parameters of the employed metasurfaces. These well-trained forward models will not only support data augmentation for expanding the training dataset but also serve as a foundation for inverse design tasks, where one-to-many mapping issues frequently arise. Next, we will construct inverse models to predict suitable metasurface configurations from desired photonic properties. Tailored to specific application scenarios, the inverse mapping will support decision-making in holographic lithography, ensuring high accuracy, speed, and robustness. Our investigations will be primarily data-driven, utilizing advanced deep learning algorithms implemented in Python. PyTorch will be the primary framework employed. Due to the high computational cost of training large-scale neural networks, as well as the need to simulate and process large datasets of photonic structures and their optical responses, the use of HPC resources is essential. Parallel computing capabilities and GPU acceleration provided by the HPC platform will be critical for model training, hyperparameter optimization, and batch simulations during the design cycle.