The influence of various training algorithms on the effectiveness of thermal conductivity prediction for MgO-GO/water–ethylene glycol hybrid nanofluid: A more effective method for network training
South African Journal of Chemical Engineering, cilt.58, 2026 (ESCI, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 58
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.sajce.2026.100968
- Dergi Adı: South African Journal of Chemical Engineering
- Derginin Tarandığı İndeksler: Emerging Sources Citation Index (ESCI), Scopus, INSPEC, Directory of Open Access Journals
- Anahtar Kelimeler: Energy conservation, Graphene oxide, Hybrid nanofluid, Performance, Thermal conductivity prediction, Training algorithms
- İstanbul Gelişim Üniversitesi Adresli: Evet
Özet
Accurate prediction of the thermal conductivity of hybrid nanofluids is essential for the design and optimization of advanced thermal management systems. In the present study, an artificial neural network framework was developed to predict the thermal conductivity of magnesium oxide–graphene oxide/water–ethylene glycol hybrid nanofluids using experimentally measured data. A total of 45 experimental datasets were generated by varying the temperature from 20 to 60 °C and the nanoparticle volume fraction from 0 to 0.20 vol.%. A feedforward multilayer perceptron network was constructed, and ten backpropagation training algorithms were systematically evaluated to identify the optimum predictive model. Among the investigated algorithms, the Levenberg–Marquardt algorithm exhibited the highest predictive performance, achieving a mean squared error of 1.976 × 10⁻⁶, a root mean square error of 1.395 × 10⁻³ W/m·K, a correlation coefficient of 0.9965, and a coefficient of determination of 0.9920. The robustness and generalization capability of the developed model were further confirmed through regression analysis, residual analysis, and 5-fold cross-validation, which yielded root mean square errors ranging from 8.38 × 10⁻⁴ to 4.73 × 10⁻³ W/m·K. A comparison with support vector regression, random forest, and Gaussian process regression demonstrated that the proposed model achieved highly competitive predictive accuracy while maintaining stable performance across different validation datasets. Furthermore, global Sobol sensitivity analysis identified nanoparticle volume fraction as the dominant governing parameter, whereas temperature exerted a considerably smaller influence on thermal conductivity. The observed enhancement in thermal conductivity was primarily attributed to the formation of conductive particle networks and improved interfacial heat transport associated with increasing nanoparticle loading. The proposed framework provided an accurate, robust, and computationally efficient methodology for predicting the thermophysical behavior of hybrid nanofluids and can be readily extended to estimate other thermophysical properties using appropriate experimental datasets.