The escalating complexity of modern network ecosystems, characterized by the integration of 5G/6G, hyperscale cloud-to-edge continuums, and massive IoT deployments, has rendered traditional trial-and-error network management obsolete. To address the need for deterministic performance in volatile environments, the concept of the Network Digital Twin (NDT) has emerged as a transformative paradigm. An NDT is a high-fidelity, real-time virtual replica of a physical network that enables continuous monitoring, \\\\\"what-if\\\\\" simulation, and closed-loop optimization. This review examines the shift toward AI-driven NDT architectures, where Artificial Intelligence (AI) and Machine Learning (ML) serve as the cognitive engine for the twin, transitioning it from a passive mirror to a proactive, predictive entity. We categorize the core architectural layers, including the data acquisition layer, the model-driven simulation layer, and the AI-powered intent-orchestration layer. The article explores how Deep Reinforcement Learning (RL) and Graph Neural Networks (GNNs) enable the NDT to perform autonomous traffic engineering, fault prediction, and security stress-testing without impacting the live production environment. Furthermore, the review addresses critical challenges such as data synchronization latency, the \\\\\"fidelity-complexity\\\\\" trade-off, and the requirement for Explainable AI (XAI) to ensure operator trust in autonomous recommendations. By synthesizing recent academic breakthroughs and industrial frameworks, this paper provides a strategic roadmap for building \\\\\"Self-Evolving Networks.\\\\\" The findings suggest that AI-driven NDTs are the foundational technology required to achieve the vision of zero-touch network management, providing a safe, intelligent sandbox for the next era of global digital infrastructure.