A physics informed neural network (PINN) framework for efficient prediction of nanofluid flow and heat transfer is built including a new inverse dynamic prediction scheme of governing parameter sets. Two PINN models based on physics-guided loss functions formulated by governing equation residuals, the Navier-Stokes equations informed neural network (NS-NN) and the Reynolds averaged Navier-Stokes equations informed neural network (RANS-NN), are presented. A standard four-module computational code is developed in C++ for comparative studies. Code validations for theoretical vortex flow fields show that the prediction errors are less than 1%, 1%, and 5% for outputs, first and second gradients, respectively. Inverse parameter sets are obtained automatically and dynamically with a root means square control mechanics for various physics governing parameters simultaneously.  Both transient flow and temperature fields for natural convective nanolfuid flow in a magnetic field are predicted with high accuracy. Moreover, enhancement of the instability is observed with the help of the PINN compared to some special cases of ground truth data.  Results show that the PINN with inverse dynamical parameters can achieve faster convergence speed and higher accuracy compared to conventional NNs.

暂无