Machine learning-assisted correlations for prediction of fission gas fractions and hydrogen concentration in VVER-1000 fuel

Abstract

This study aims to develop correlations for predicting the fission gas fractions and hydrogen gas concentration during a fuel cycle, using gene expression programming as an evolutionary machine learning approach. The well-known FRAPCON code is used for generating a straightforward dataset under steady-state conditions. The two-step sensitivity analysis is carried out to identify the most influential parameters for correlation development. Wilks’ statistical method is used to generate 59 scenarios to distribute input parameter uncertainties evenly, which leads to a confidence level of 95 %. The mean squared error for xenon, krypton, and helium is 0, while hydrogen exhibited a value of 59.36 since fraction values are between 0 and 1 and concentration ranged from 5 PPM to 200 PPM. $R^2$ values exceeded 0.97, indicating strong correlation accuracy. The high accuracy achieved from the correlations demonstrates that selecting a 59-sample dataset based on Wilk’s method is sufficient to obtain accuracy exceeding 95 %.