Machine Learning Based Surrogate Modelling for Accelerating CRWEA Analysis : A Residual CNN Approach for Predicting Reactor State

Sammanfattning

Safety analysis during the nuclear reactor core design process is computationally demanding due to the complexity of the underlying physics based simulations. For boiling water reactors, an analysis on Control Rod Withdrawal Error Accident (CRWEA) is made to meet safety demands on core design. The analysis relies on nuclear reactor simulation codes which deterministically solves the neutron diffusion equation in a discrete space. Multiple of these simulations are done throughout the analysis which result in a significant evaluation time. This motivates the investigation of surrogate machine learning models capable of reducing the computational cost while maintaining acceptable prediction accuracy. In this project, a surrogate machine learning model for predicting the 10x10 minimum Critical Power Ratio (CPR) distribution around a control rod at a position during a burnup step was implemented. This thesis was conducted at Vattenfall Nuclear Fuel and the aim was to investigate the feasibility of replacing computationally expensive nuclear reactor simulations calculations during CRWEA analysis with a machine learning based approach. To explore this feasibility, a convolutional neural network with residual blocks was implemented and trained on reactor simulation data generated using the nuclear reactor simulations software. To account for the safety critical importance of lower CPR values, a weighted asymmetric mean squared error loss function was implemented during training. The results indicate that the implemented approach is capable of predicting the minimum CPR distributions with promising accuracy while significantly reducing the computational time compared to direct nuclear reactor simulations calculations. These results suggest that surrogate machine learning models could potentially be used to accelerate safety analysis during reactor core design process.