Enhancing Nuclear Safety Margins: A Hybrid Layered Architecture Combining Deterministic Physics and Machine Learning

The safety of nuclear reactors critically depends on the early detection of thermohydraulic and neutronic anomalies. Traditional Reactor Protection Systems (RPS) are deterministic and effective for accident mitigation, but limited in identifying incipient faults (drifts) and prone to generating false alarms in noisy regimes, leading to alarm fatigue. This work proposes a Physics-Informed Digital Twin architecture that combines integrated Neutron Point Kinetics via 4th Order Runge-Kutta (RK4) with a One-Class Support Vector Machine (OC-SVM) classifier. Unlike purely data-driven ("black box") approaches, the AI analyzes only the stochastic residuals between the physical prediction and the measurement, ensuring interpretability and physical conformity. Simulated results demonstrate that the system is capable of detecting a loss of cooling efficiency approximately 800 seconds before the violation of classic safety limits, while maintaining immunity to false positives through a temporal persistence filter.