Perpustakaan Digital ITB

Despite its well-known ability in solving a fluid flow, the CFD method has several shortcomings, including a rigorous meshing process and careful numerical method consideration, which inevitably results in a trade-off between accuracy and computation time. In this thesis, we have an interest in tackling these shortcomings by employing a machine learning method called physics-informed neural network (PINN). In its early study, PINN shows capability in solving several fluid flow problems. However, it is still in its infancy. To assess PINN potential as an alternative flow field solver, this thesis studies the PINN capability in terms of its formulations, hyperparameters, and configurations on several benchmark flow problems that focus on accuracy and computation times metrics. Moreover, we also performed an in-depth study about the transfer learning method that could accelerate the PINN computation time, whereby is favorable specifically in design exploration. The results concluded that PINN performed well in solving fluid flow by producing a good accuracy on the benchmark problems. Furthermore, the transfer learning method significantly reduces PINN computation time, even when engaged in different domain and boundary conditions.