A hybrid simulation method integrating CFD and deep learning for gas-liquid bubbly flow

This paper develops a hybrid framework that integrates deep learning and CFD simulation for physical field prediction of bubbly flow. The framework utilizes PyTorch for constructing the deep learning model, and OpenFOAM for conducting the CFD simulations. In recursive predictions, i.e. predicting the flow field by the deep learning model successively, the prediction error increases. In contrast, the hybrid simulation mitigates error accumulation and can run for long flow time (T = 300 s) without divergence. Besides, the hybrid simulation effectively predicts the variation in bubble quantity within the calculation domain. The velocity fluctuation at the measurement point can be qualitatively captured, with reasonable fluctuation amplitude close to the prediction of Euler-Lagrange solver DPMFoam. However, there is still some error in the fluctuation period, which may be further reduced by improving the deep learning model or shortening its prediction time span. In the studied case, the hybrid simulation saves a total of 40% of the computation time compared with OpenFOAM simulation. This work provides a feasible way to combine deep learning and CFD for studies of the gas-liquid bubbly flow and more related multiphase flows, and makes contribution to the long-term accelerated simulation.