Model-Free Solution Estimation of Critical Heat Flux Using Genetic Programming

Many equations in thermal hydraulics used to predict critical heat flux (CHF) stem from purely empirical data or semimechanistic models, with the latter inherently relying on a priori theoretical assumptions and expert intuition. These correlations must ensure not only accurate predictions but also physical soundness, particularly in terms of dimensional consistency. In this study, a model-free solution estimation approach using genetic programming (GP), an easily explainable artificial intelligence model, is investigated to provide alternatives to the regression process. Without relying on prior mechanistic assumptions, GP extracts symbolic correlations directly from data, while preserving dimensional consistency. In a multivariate analysis, four groups of databases were used: (1) an initial set generated using Kandlikar’s model, (2) an extended multisized set derived from the initial one, (3) a noise-based set based on the initial two sets, and (4) an experimental dataset. For the most challenging dataset, the last one, composed of a fraction of the experimental data Kandlikar used to develop his correlation, predictions remained in a 30% error region, explained by the low number of data points to train the model. In general, GP performance was quite acceptable, demonstrating a unit-aware symbolic regression workflow that discovered explicit CHF correlations directly from dimensional variables, for all four databases. GP, a powerful tool for future applications, is demonstrated, aiming to help in the process of obtaining new and unknown correlations. The wide range of the model’s applicability for different fields of study is also emphasized, with easily interpretable results.