Chaotic Sech-Tanh dynamic opposition-based learning for metaheuristic optimization. Part II: deep neural network architecture search for two-phase pressure drop prediction

Abstract

Accurate prediction of two-phase pressure drop is essential for the design of the tubular heat-exchange equipment used throughout energy-conversion and thermal-hydraulic systems, yet classical empirical correlations generalize poorly across fluids and operating conditions. This study develops a metaheuristic-driven neural architecture search for predicting the pressure drop due to evaporation and condensation of refrigerants in smooth and micro-fin tubes. Eight state-of-the-art metaheuristic optimizers, each enhanced with the Chaotic Sech–Tanh Dynamic Opposition-based Learning strategy developed in a companion study, were used to optimize the architecture and hyperparameters of a deep neural network trained on an experimental database, comprising 946 measurements (626 evaporation and 320 condensation points). The optimized networks were benchmarked against nine established machine learning regression models. A network with three hidden layers of 50 neurons achieved a mean absolute error (MAE) of 0.157 and a coefficient of determination ( R ² ) of 0.99446 on the evaporation data, with 88.6 % of predictions falling within the 5 % error band. The opposition-enhanced optimizers consistently outperformed their standard counterparts: the improved Marine Predators variant was the most accurate for evaporation (MAE 0.224, R ² 0.98977), and the improved Mantis Search variant was the most accurate for condensation (MAE 0.957, R ² 0.98347). Among the machine learning models, extreme gradient boosting attained the highest training accuracy ( R ² up to 0.99826) but showed signs of overfitting, whereas random forests generalized best.