Study on application of artificial neural network to debris bed coolability calculations for sodium-cooled fast reactors

Abstract

Understanding the effect of uncertainties of Core Disruptive Accident (CDA) scenarios on debris bed coolability on a core catcher is required for decision making on design options to mitigate a CDA consequence. For the understanding, a huge number of calculations are required but are extremely difficult to perform because a huge number of calculations require much calculation time to solve non-steady equations in the coolability calculation model. Thus, we applied Artificial Neural Network (ANN), which is one of models for machine learning, to debris bed coolability calculations. The application of ANN is expected to exponentially improve the calculation speed of debris bed coolability because ANN provides results from experimental rules learned through training without solving non-steady equations. The application is in three steps. Firstly, we created many data for training ANN and validating the trained ANN through coolability calculations parameterizing main dominant inputs (particle diameter of debris bed, porosity of debris bed, etc.) by using Latin hypercube sampling. Secondly, ANN was trained and validated with the created data. The accuracy rate of the results by the ANN to the validation data exceeded 99%. In addition, the calculation time using ANN was micro seconds order. Finally, through demonstration calculations, it was confirmed that we can easily understand the effect of uncertainties of CDA scenarios on debris bed coolability owing to results visualization based on a huge number of parametric calculations using ANN. Thus, the application of ANN to debris bed coolability calculations should contribute to the decision making on design options to mitigate a CDA consequence.

Keywords

• sodium-cooled fast reactor
• debris bed coolability
• speed-up calculations
• artificial neural network
• machine learning
• application
• latin hypercube sampling