Experimental and multioutput explainable machine learning investigation of thermal-hydraulic-entropic analyses in channels equipped with sinusoidal turbulators

This study investigates the thermo-hydraulic and entropic behavior of air flow in channels equipped with sinusoidal turbulators through a combined experimental and machine learning framework. Unlike previous studies that focused on conventional thermohydraulic measurements or single-output black-box machine learning predictions, this study presents the first experimentally validated scope of multi-output explainable machine learning for the simultaneous prediction of heat transfer, pressure loss, and entropy generation. Experiments were conducted for Reynolds numbers between 17,000 and 73,000 and three turbulator widths (a = D/4, D/2, 3D/4) to assess their effects on heat transfer, pressure drop, and entropy generation characteristics relative to a smooth reference channel. The smallest turbulator width provided the most balanced thermo-hydraulic behavior within the investigated range and yielded a substantial reduction of about 47% in total entropy generation compared to wider turbulators. On the otherhand, new correlations have developed for Nusselt number, Darcy friction factor, and entropy generation according to experimental findings. A multi-output machine learning pipeline was developed in Python to predict seven key performance indicators from 32 experimental samples using physics-informed feature sets and both linear and nonlinear regressors. Explainable AI analysis using Shapley Additive exPlanations identified Reynolds number and turbulator width (a/D) as the dominant factors and provided an interpretable mapping between operating conditions, geometry, and second-law behavior. The proposed hybrid framework offers an accurate and interpretable basis for within-range performance prediction and design comparison of enhanced heat-transfer channels.