Emulation of cardiac mechanics using Graph Neural Networks
David R. Dalton, Hao Gao, Dirk Husmeier
Abstract
Recent progress in Graph Neural Networks (GNNs) has allowed the creation of new methods for surrogate modelling, or emulation, of complex physical systems to a high level of fidelity. The success of such methods has yet to be explored however in the context of soft-tissue mechanics, an area of research which has itself seen substantial developments in recent years. The present work explicates on this by introducing an emulation framework based on a multi-scale, message-passing GNN, before applying it to the modelling of passive left-ventricle mechanics. Through numerical experiments, it is demonstrated that the proposed method delivers strong predictive accuracy when benchmarked against the results of the nonlinear finite-element method (FEM), and significantly outperforms an alternative emulator based on a fully connected neural network. Furthermore, large computational gains are achieved at prediction time against the FEM.