ANN-Based Prediction of Thermo-Hydraulic Performance of Mgo and Mgo–Cuo Transformer Oil Nanofluids in a Plate-Fin Heat Exchanger

Abstract

Accurate prediction of the coupled thermal and hydraulic behaviors of nanofluids in heat exchangers potentially avoids repetitive experimental testing. Thus, the present study aimed to develop a data-driven surrogate modeling framework capable of predicting the thermal and hydraulic behaviors of MgO and MgO-CuO / Transformer Oil nanofluids in a plate-fin heat exchanger in a temperature range of 30–70 ◦C. We digitized literature-sourced data using WebPlotDigitizer and reconstructed heat transfer and flow-rate experimental curves with piecewise cubic Hermite interpolation, maintaining monotonic behavior in the time dimension. Four independent forward feedforward 2–5–1 architecture artificial multilayer perceptrons, trained using the Levenberg-Marquardt technique, predicted overall and convective heat transfer coefficients, Reynolds number, and pumping power. Reliability was demonstrated through 30 individual runs, 10-fold cross-validation, repeat internal holdout, grouped anchor holdout, and boundary stress tests. The results were compared with support vector regression with radial basis function (SVR-RBF), random forest regression (RFR), and a multilayer perceptron trained via stochastic gradient descent with momentum. The overall percentage error (mean absolute error, MAE) for the four parameters for the Levenberg-Marquardt approach was 2.363, 0.433, 1.055, and 0.280%; the mean coefficient of determination (R2) ranged from 0.97838 to 0.99789. No approach dominated overall; alternative models had minimal training error for individual output variables, while the LevenbergMarquardt framework produced the most evenly distributed results and the lowest variability among the four coupled responses. Under boundary evaluation, this framework captured the highest overall efficiency (lowest root mean square errors for both the convective heat transfer coefficient, 25.491 W/m²⋅K, and the Reynolds number, 21.653). Under the given boundary conditions, changes in oil viscosity due to heat input altered hydraulic flow-rate behavior, while adding nanoparticles changed thermal efficiency and the pumping energy cost ratio. By integrating multiple response models, comparing algorithms, and using a boundary-validation approach, the constructed framework may provide very fast intra-domain filtering of the thermo-hydraulic design. Nevertheless, its usability is limited to the described test conditions for the mentioned fluid and exchanger type.