Efficient training of machine learning models using non-converged iterates in nuclear reactor simulations

Machine learning (ML) algorithms exhibit promising potential for enhancing safety, improving predictive accuracy, and streamlining nuclear reactor system analysis, an opportunity that is currently the focus of intensive research efforts. However, gathering reliable datasets for training is often computationally expensive and time-intensive, making ML model development impractical in some cases. A novel methodology is introduced in this work to solve this issue efficiently by utilizing non-converged iterations from iterative nuclear reactor simulators. Two case studies are presented to demonstrate and test the proposed method. In the first case study, an artificial neural network (ANN) model is trained to establish the relationship between group flux and the multiplication factor using a dataset collected from non-converged iterations of the neutronics solver. Similarly, in the second case study, a support vector regression (SVR) model is constructed and trained to capture the effect of fuel gap conductivity on the system's multiplication factor. The results demonstrate that training ML models with non-converged iterations can produce reliable predictions of the quantities of interest with acceptable accuracy. For case studies 1 and 2, root mean square errors of 148 pcm and 10 pcm were achieved, respectively. This approach also led to substantial computational cost reductions, exceeding 98% savings in data acquisition cost (simulation runtime) in both cases. Furthermore, a comparative analysis with ML models trained on fully converged solutions revealed that while the accuracy of the latter was superior, the proposed approach yielded acceptable predictive performance at a fraction of the computational cost, underscoring its practicality for scenarios where computational resources are constrained.