Adaptive Sampling Framework for Data-Efficient AI Prediction of Parameter Trends in Small Modular Reactor Simulations

Small modular reactors (SMRs) require reliable monitoring and operator-support capabilities during abnormal conditions, but AI forecasting models for SMR transients require simulation data spanning malfunction severity and mitigation timing. This study proposes a hierarchical adaptive sampling framework for data-efficient generation of forecasting datasets in the IAEA integral pressurised water reactor (iPWR) simulator. Level 1 adaptively selects leakage severities from a no-action pool to diversify transient regimes and reduce redundant sampling in saturated early-trip regions. Level 2 refines isolation-valve closure timing for selected severities to densify samples near trip-margin-relevant timing bands. An LSTM model is iteratively trained on accumulated scenarios and used to score informative candidates by forecasting error. Results for a pressurizer relief-valve leakage case show that the method concentrates samples near severity and timing transition boundaries, improving regime coverage under a limited simulation budget. The framework is intended as an offline simulation-campaign strategy supporting future AIbased predictive monitoring and operator-support applications.

Keywords: Adaptive sampling, small modular reactor, iPWR simulator, parameter trend prediction, operator support, LSTM.