Physics driven interpretable deep learning-based insights into boiling crisis of smooth and roughened surfaces

The safety and reliability of phase change equipment in many applications rely on accurate prediction of pool boiling critical heat flux (CHF). The existing predictive models (empirical correlations or some semi-analytically based theoretical approaches) generally predict the CHF with very large errors owing to the vague pool boiling phenomenon. Herein, a physics informed and interpretable artificial intelligence based predictive model was developed to accurately predict the CHF using deep neural network (DNN) for a range of smooth and roughened surfaces with diverse working fluids and pool boiling conditions. An automatic k-fold method known as Bayesian surrogate models (Gaussian process, Gradient boost regression trees, Random forest) is used to provide an optimal and stable model while incorporating big database regarding CHF. Through empirical correlations developed in 42 pool boiling studies incorporated with 29 working fluids on various surfaces and correlation matrix or data-based correlations such as Pearson correlation, Kendall correlation, and Spearman correlation, we have identified the key input parameters based on surface morphologies, pool boiling conditions, and thermophysical features to predict CHF (the accuracy achieved R² = 0.97). Moreover, the interpretable artificial intelligence (XAI) has been used to shed a light on the DNN's predictions. The XAI reveals that surface inclination is the most influential parameter followed by liquid saturation temperature, material conductivity, ΔT, and surface roughness for the investigated data range. The parameter sensitivity analysis strongly agrees with XAI.