Abstract: In recent years, risk-informed safety margin characterization (RISMC) methodology has emerged as an advanced tool for probabilistic nuclear safety assessment, which combines the advantages of deterministic safety analysis and probabilistic risk assessment. RISMC enables dynamic quantification of accident risk and provides technical support for risk-informed decision-making during plant operation. However, the practical application of RISMC faces significant challenges, including low computational efficiency, high demand for computing resources, and limited capability for real-time risk prediction, especially under complex and rapidly changing operating conditions in modern nuclear power plants. These limitations restrict the ability of traditional methods to provide rapid and accurate safety margin evaluations, which are essential for timely operator response and effective accident mitigation. To address these limitations, this study proposed an efficient safety margin evaluation framework that integrates high-fidelity thermal-hydraulic (T-H) simulations with a data-driven surrogate model. Specifically, T-H simulations were used to generate large-scale response datasets representing different accident scenarios and operational strategies. Based on this dataset, a Transformer-based surrogate model was trained, which leverages self-attention mechanism to effectively capture complex nonlinear relationships and global dependencies among high-dimensional system parameters. Unlike traditional machine learning models, such as support vector machine (SVM) and multi-layer perceptron (MLP), the Transformer architecture offers significant advantages in modeling dynamic sequences and handling high-dimensional, strongly coupled system features. The proposed framework significantly improves both computational efficiency and prediction accuracy in safety margin analysis, enabling rapid evaluation of reactor safety margins under a variety of operator intervention conditions. To validate the effectiveness and robustness of the proposed method, a comprehensive case study was conducted focusing on the partial loss of feedwater (PLOFW) accident in the secondary circuit of a typical pressurized water reactor nuclear power plant. The performance of the Transformer-based surrogate model was quantitatively compared with conventional regression models, including SVM and MLP, using metrics such as mean squared error (MSE) and the coefficient of determination (R²). The results demonstrate that the Transformer-based model reduces the MSE by approximately 90%, improves the R² value to 0.982, and achieves reliable and rapid predictions of critical safety parameter limits under various operator intervention strategies. Furthermore, by efficiently quantifying the probability of exceeding safety limits and reactor shutdown risk, the proposed approach enhances operators’ situational awareness, supports real-time monitoring, and provides timely, quantitative, and risk-informed decision support for accident mitigation and safe operation of nuclear power plants. This work demonstrates the great potential of integrating advanced surrogate modeling with RISMC methodology for improving the flexibility, efficiency, and accuracy of nuclear safety margin management in complex engineering environments.