Comparative Analysis of Sensor Configurations for Blood Flow Predictions via Physics-Informed Neural Networks

This study presents a comprehensive investigation into the impact of sensor configurations on the accuracy of flow-field predictions by a Physics-Informed Neural Network (PINN) model in the context of a 2-dimensional stenosis hemodynamics problem. Utilizing the Latin Hypercube Sampling (LHS) technique, we systematically generated 80 distinct sensor configurations across the computational domain. Our results demonstrate that the accuracy of flow-field predictions is notably more sensitive to sensors located close to the stenosis and inlet, challenging the conventional assumption of equal importance for all sensors. The spatial arrangement of sensors emerges as a critical factor influencing predictive accuracy. The sensor configurations with the highest and lowest prediction accuracy had Root Mean Square Error (RMSE) values of 0.3% and 30%. These findings offer valuable insights for the optimization of sensor layouts in hemodynamic modeling, indicating that careful consideration of the configuration, especially in regions near the stenosis and inlet, can significantly enhance the accuracy of flowfield predictions. This research contributes to the refinement of sensor deployment methodologies and advances our understanding of the relationship between sensor placement and predictive precision in stenosis hemodynamics modeling using PINNs.