Neural Network-Based Surrogate Models Applied to Fluid-Structure Interaction Problems

Traditional computational methods face significant challenges with ever-increasing complexity in the problems of engineering interest. One category of problems that suffer from this phenomenon is those where Fluid-Structure Interaction (FSI) is present. One set of problems that suffer from this phenomenon is those where Fluid-Structure Interaction (FSI) is present. FSI simulations are traditionally time-consuming and computationally extremely expensive. Potential alternatives rely on using a surrogate model to substitute one or more systems involved. A promising approach employs artificial neural networks as the basis for such a surrogate model combined with strong physics simulations based on finite element methods (FEM). This approach requires the seamless integration of AI algorithms and packages into the simulation workflow. Such an example is the NeuralNetworkApplication developed in KratosMultiphysics. The routines related to the neural networks are executed through an interface with the Keras API. Mok's benchmark is chosen as the study case to test the capacity of the previous method applied to FSI problems. Two cases in which one of the systems is substituted by a neural network-based surrogate model are analyzed. Strong and weak coupling scenarios are considered. The results present improvements in simulation time without sacrificing accuracy, especially when compared with the original benchmark. This contribution discusses the influence of the original data and network architecture on the simulation outcome and different considerations for generating surrogate models for FSI.