AI-enabled cardiovascular models trained on multifidelity simulations data

Neural networks (NNs) are routinely used to accelerate ensemble-based computations (EBCs) of physical systems governed by partial differential equations. However, the high computational costs of training NN-based surrogates rival those of EBCs, limiting their practical utility. We address this challenge by integrating multifidelity data, obtained by running models of different complexity and reliability, and transfer learning to optimize the training of deep convolutional neural networks (CNNs). The data are generated via patient-specific hemodynamic simulations of an aorta, leveraging high- and low-fidelity models and systematically varying boundary conditions and blood properties within the Monte Carlo framework. Our results demonstrate that CNN surrogates benefit from a preliminary tuning on low-fidelity data and achieve high accuracy, in comparison to highfidelity simulations, with significantly reduced computational costs, paving the way for efficient, patient-specific cardiovascular assessments.