Wind Farm Layout Optimization (WFLO) is a critical process in wind farm design aimed at determining the optimal placement of turbines to maximize annual energy production (AEP). Accurate AEP estimation in WFLO depends on the precise prediction of the flow field within the wind farm. This is typically achieved using either analytical wake models or computational fluid dynamics (CFD) simulations. Analytical wake models approximate the wake flow by relying on simplified representations of the governing physical laws, making them computationally efficient but less accurate in capturing complex flow structures. In contrast, CFD simulations provide accurate flow fields by solving the Navier–Stokes equations for turbulent flow across the entire wind farm. However, their high computational cost makes them impractical for large-scale optimization studies. This thesis presents the development of a deep convolutional hierarchical encoder–decoder neural network, DeepWFLO, as a low-cost surrogate model for predicting flow fields in wind farms, thereby accelerating the WFLO process. DeepWFLO is capable of predicting flow fields for various layouts by capturing the wake interactions among multiple turbines. To enhance physical consistency, this work proposes a physics-informed variant, PI-DeepWFLO, which incorporates the Reynolds-Averaged Navier–Stokes (RANS) equations into a custom loss function. Comparative analyses demonstrate that PI-DeepWFLO achieves improved stability and robustness by reducing sensitivity to variations in network weight initialization and training datasets relative to the purely data-driven DeepWFLO. Furthermore, an active learning framework is introduced to adaptively select the most informative training samples, thus reducing the amount of labeled data required to achieve a target accuracy. Overall, these contributions provide a computationally efficient and accurate surrogate model for flow field prediction in wind farms, enabling more effective and scalable WFLO.