A soft computing approach for estimating the specific heat capacity of molten salt-based nanofluids

Despite the promising potential of nanofluids as heat transfer and energy storage media, determination of their thermal behavior and properties need significant experimentation. Considering the relatively high costs of such fluids and the time-consuming procedures for synthesizing them and measuring their characteristics, machine learning techniques can be powerful tools for simulating their behaviors in the unstudied combinations of operating conditions. In this study, a machine learning model has been developed for the first time in the literature - to simulate and predict the specific heat capacity of a molten nitrate salt mixture seeded with silica, alumina and titania nanoparticles. A multilayer perceptron neural network (ANN) was selected among 1920 ANNs with different architectural features. With a prediction R-2 value of 0.9998, the suggested model was found to provide much superior predictions (and validated against experimental data) as compared to the classic analytical models. The model developed in this study can, therefore, be used for estimating the values of specific heat capacity for nanofluid samples - based on the temperature and mass fraction of the nanoparticles, as well as the average (or nominal size) of the nanoparticles. The soft-computing technique itself was evaluated under extreme training conditions and it was found that the algorithm can adapt to new data sets with maximum MAPE of 2% and can enable excellent quality of predictions (R-2 > 0.95) when trained with <300 data points. (C) 2019 Elsevier B.V. All rights reserved.