Super-resolution analysis: reducing computational cost of high fidelity simulation of flashing flows.

High fidelity numerical simulation incur increased computational cost when directly solving partial differential equations (PDE) using traditional techniques like finite volume method (FVM). With the recent advancement in physics based machine learning, the available flow data  in the coarse grids could help in reconstructing solutions of higher fidelity. In this study, we use the super-resolution analysis technique to obtain higher resolution results of a flash boiling flow from an initial coarse grid simulation. A standard convergent-divergent nozzle as a domain  in  this supervised neural network study. We employ down-sampled skip-connection multi-scale convolutional neural networks to reconstruct solutions on a fine grid. Stochastic gradient descent technique was implemented to update the weights by calculating the loss function.  To maintain the size of the output after  the convolution,  reflective padding was used. The effect of different kernel sizes,  learning rates and hidden layers were also investigated in this study. The results are obtained with reasonable accuracy and at a fraction of computational cost as compared with a conventional numerical  method to solve PDE.