Transient velocity and temperature field predictions for natural convective nanofluid flows immersed in oscillating magnetic fields with physics informed neural networks

A physics informed neural network (PINN) framework is constructed for predicting the transient fields of natural convective nanofluids immersed in oscillating magnetic fields. A PINN code with inverse inference of dynamic parameters is developed. For pure fluid without nanoparticles, prediction errors of temperature and velocity fields by the PINN are 0.5% to 1.1% smaller than the conventional neural network (NN). Across all training and validation datasets for both pure fluid and nanofluid cases, the convergence speed of the PINN is about 10% faster than that of the NN at 90% accuracy level. Due to combined effects of parameters on multiple inverse coefficients, the recovered nanoparticle volume fractions of nanofluid properties and Hartmann numbers of magnetic field intensities exhibit larger inference errors of 1.5% and 3.0%, respectively, compared to the two transport parameters, i.e., Rayleigh numbers and Prandtl numbers with inference errors less than 1.0%. The convergence speed of the nanoparticle volume fraction is considerably slower than that of all other parameters, because it affects all effective properties and governing parameters. Investigations with Hartmann numbers ranging from 0 to 1000 show varying performance in different heat transfer regimes. In the convection dominating range with Hartmann numbers less than 100, the PINN significantly outperforms the NN, achieving reductions at least 4.81% and 4.21% in the testing datasets for velocity and temperature fields, respectively. The PINN improves the fidelity of forecasting and inverse inference, particularly for stronger convective nanofluid flows.