This thesis applies machine learning (ML) methods to the numerical simulation of a fluid-structure interaction problem involving an elastic tube containing a liquid. A neural network, utilizing fully connected and recurrent layers, is trained on the simulation data such that the prediction during inference generates the subsequent time step to a given input batch. To make use of the parallelization technique presented in [1], we partition the tube and train a separate neural network on each partition independently. Due to low accuracy and continuity of poor quality in the predictions, we further make use of having adjacent partitions overlap such that information can be shared.

Our results suggest using multiple layers in the neural network architecture is superior to only having single layers. Additionally, the application of L2 regularization in form of weight decay has no detrimental effect on the error of the predictions. The overlapping technique demonstrates a highly valuable method of increasing the continuity of the predictions whereas we find no significant difference regarding the amount of overlap used. Finally, the implemented parallelization approach can make better use of machines providing a higher number of CPU cores when compared to the multithreading offered by the used ML library directly.

[1] Totounferoush, Amin, et al. "Parallel Machine Learning of Partial Differential Equations." 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW). IEEE, 2021.