This review examines the transformatively expanding contribution of Artificial Intelligence (AI) to fusion science. It focusses on machine learning (ML) and deep learning (DL) as foundations of modeling, control, and comprehension of data. It evaluates the mechanism by which AI raises predictive capability, efficient computation, and scientific comprehension within the fusion workflow, while critically examining limitations keeping full realization elusive. By conceptual research, comparative modeling, schematic infrastructure, ablation experiments, and hybrid methodologies such as Physics-Informed Neural Networks (PINNs), the article examines the interplay between AI and the rich data environments of fusion. A survey of the last decade's peer-reviewed publications reveals that ML and DL enable up to 10× faster diagnostic inference, reinforcement learning achieves real-time plasma control making thousands of adjustments per second, PINNs reduce transport model computation by 5× while cutting cost by 72%, and AI–physics hybrid modeling raises predictive accuracy to 74% while surpassing conventional simulation. Despite all of these, challenges persist. The fusion data remains diverse, resistant to standardisation, lack of interpretability is a common failing among the models, dynamic reactor scenarios demand recurrent recalibration, the restrictions around ethics, operations, and collaboration complicate roll-out. This review concludes that AI must become physics-aware, adaptive, and transparent. By embracing the domain expertise and facilitating the federated learning bases, AI becomes complementary not a replacement to the traditional scientific method, thereby offering the future path to sustainable energy innovation.
Emmanuel et al. (Tue,) studied this question.