AI across the nuclear lifecycle: moving from isolated demonstrations to trustworthy decision support requires assessing complete workflows, not just predictive models

The review organizes AI applications in nuclear engineering into an evidence map spanning the lifecycle, including nuclear datasets and computational infrastructure, surrogate and physics-informed modeling, digital twins, monitoring and prognostics, lifecycle applications, optimization and control, and trustworthy AI. Relative to studies focused on a single demonstration or model, this work groups dispersed reports by lifecycle stage and capability type, letting readers see the overall scope in which AI is being investigated in the nuclear domain. An inductive review of reported studies; evidence is presented through thematic categorization rather than a unified quantitative meta-analysis.

The review distills three reportable capabilities of learning-based methods in nuclear engineering: accelerating high-fidelity calculations, extracting information from multivariate measurements, and supporting decisions in reactor operation, maintenance, waste management, and environmental assessment. It consolidates capability statements previously scattered across application contexts into three comparable functions, helping distinguish where AI has been investigated from where it remains exploratory. A synthesis of cited reported studies; the text provides no unified effect sizes or performance aggregation across studies.

The review identifies recurring limitations: scarce abnormal-condition data, differences between simulated and physical systems, uncertain generalization, and incomplete evaluation of downstream decisions. It presents these limitations as patterns recurring across applications rather than isolated issues in individual studies, shifting the discussion from model accuracy toward deployment conditions. A qualitative synthesis of commonalities across multiple application reports, not a controlled comparison.

The review argues that progress depends on assessing complete AI-enabled workflows rather than predictive models alone, and that representative data, independent testing, calibrated uncertainty, physical and operational constraints, model traceability, and human oversight are especially important when outputs affect safety-relevant decisions. It expands the unit of assessment from the model to the end-to-end workflow and lists concrete requirements matched to safety-relevant decisions, setting an evaluation frame for subsequent research. An argumentative claim advanced by the authors on the basis of the reviewed evidence, supported by discussion rather than experiment.

The work addresses researchers, system designers, and evaluators of safety-relevant decisions in nuclear engineering, and applies to settings where one must judge whether an AI demonstration is operationally usable. Its proposed evaluation points—representative data, independent testing, calibrated uncertainty, physical and operational constraints, model traceability, and human oversight—target stages whose outputs affect safety-relevant decisions. The review also states that future research should connect simulation, sensing, digital twins, and uncertainty-aware decision methods while testing performance under realistic changes in operating conditions, pointing toward end-to-end validation pipelines.

As a discursive review, the text provides no unified quantitative comparison, so the relative maturity of capabilities across applications still requires readers to consult the original studies. Scarce abnormal-condition data and simulation-to-physical differences are listed as recurring limitations, but the text does not rank their severity across application types. In addition, the reviewed reports are largely presented as demonstrations, and their behavior under realistic changes in operating conditions remains an open question—which is precisely the direction the authors propose for future testing.