AI-Driven Thermal-Hydraulic Optimization of a Serpentine Water–Air Heat Exchanger Using Computational Fluid Dynamics Analysis

Serpentine water-air heat exchangers play a central role in heating, ventilation, and airconditioning (HVAC) equipment, automotive radiators, fuel-cell humidifiers, photovoltaic-thermal collectors, and the cold-plate cooling of high-power electronics; yet their high-dimensional design space-spanning hydraulic diameter, bend curvature, pass count, fin geometry, and dual-stream flow rates-defies conventional trial-and-error tuning.This study introduces a reproducible Artificial Intelligence-Computational Fluid Dynamics (AI-CFD) pipeline that integrates four elements into an end-to-end framework: a Deancorrected validated CFD-correlation engine; a head-to-head benchmark of three surrogate families on five thermal-hydraulic targets; a Non-dominated Sorting Genetic Algorithm II (NSGA-II) Pareto search; and a Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) compromise rule that returns an engineering-realistic optimum rather than a mathematical extremum.Six hundred Latin-Hypercube design points feed the surrogates.The best surrogate (a 64-64-32 multilayer Artificial Neural Network (ANN)) reaches R² = 0.989 for heat duty, 0.979 for effectiveness, and 0.987 for the Performance Evaluation Criterion (PEC), with mean absolute percentage errors below 4% for these three targets.Compared with a representative industrial baseline, the AI-selected design simultaneously raises duty by 31.2%, lifts effectiveness from 0.37 to 0.77 (≈ +108%), trims water-side and air-side pressure drops by 70.7% and 82.9%, and reduces total pumping power by 88.0%, yielding a PEC of 5.94.Sensitivity analysis reveals that the pass count, water mass-flow rate, and fin height are the dominant parameters for heat duty and effectiveness, whereas the bend radius and fin height govern the PEC through the balance between Dean-vortex enhancement and the associated friction penalty.The pipeline takes less than 90 s to complete on a workstation, which is over two orders of magnitude faster than similar direct-CFD optimization pipelines.