Towards Trustworthy Autonomy in Advanced Nuclear Systems: A Review of Advanced Control, Digital Twins, and Generative AI

Advanced nuclear systems are expected to operate across a wider range of missions, plant configurations, staffing models, and external couplings than the existing light-water reactor fleet. The operating practices of that fleet were developed for large single units at steady baseload and cannot be assumed to transfer unchanged to flexible, multi-module, remotely supervised, or multi-product plants. This review examines how advanced control, digital twins, and generative artificial intelligence (GenAI) can extend those practices toward trustworthy autonomy. Model-based and predictive control, learning-enabled control, control-oriented digital twins, supervisory operation, and GenAI grounded in validated engineering tools are synthesized within a single framework organized by level of automation. The framework considers plant layer, degree of autonomy, and evidence maturity from conceptual proposals to operating-plant experience. The review finds that higher autonomy is most credible within a layered architecture. Independent protection and qualified regulatory control remain the safety foundation that every advanced layer must defer to. Predictive and supervisory control coordinate constrain plant-wide operation, and reinforcement learning is best confined to offline-trained supervisory decisions screened by verified filters. Digital twins supply state estimation, prediction, and controller-development infrastructure whose role narrows from action to advice as uncertainty grows. GenAI is best introduced as a grounded interface that retrieves controlled information, invokes validated tools, and prepares recommendations for human approval. The principal gap is integrated demonstration, and a staged roadmap in which the level of automation rises only with accumulated evidence is offered from the author’s perspective.