Systematic process to determine DNBR limit of CHF correlation with repetitive cross-validation technique

The concerns have been raised about the potential for overfitting, which means that the correlation can predict the data used to develop the correlation well, but lacks in predictive capability on other data not used in the development of the correlation. In developing the CHF correlations to avoid overfitting problem when validation data are not enough or has a meaningless range, currently, it has been suggested that the database for a CHF correlation should be divided into a training data-set and a validation data-set. The systematic process to develop CHF correlation with cross-validation technique and to yield 95/95 DNBR values was developed to estimate quantitatively the risk of its overfitting on the resulting DNBR limit of CHF correlation. The repeated hold-out method out of cross-validation techniques was applied to the example CHF correlation (KCE-1M) by running 1000 random-sampling trials of CHF database with training (75%) and validation (25%) data-set. The effect of cross-validation technique on determination of the DNBR limit was estimated less than 3% at 95% probability with 95% confidence. Also, it was verified that DNBR limit of CHF correlation with cross-validation was more conservative to be applicable to the thermal-hydraulic design than that without cross-validation.