Outlier-resistant physics-informed neural network

Recent advances in machine learning have introduced physics-informed neural networks (PINN) as a valuable tool for addressing dynamics through governing equations and experimental observations. Outliers can be present in measurements and significantly affect the accuracy of the solutions provided by PINN. To overcome this limitation, we construct an outlier-resistant PINN (OrPINN) based on Tsallis statistics. We investigate the robustness of OrPINN in describing the acoustic and linear elastic wave dynamics under various outlier-level scenarios. We find that the OrPINN can improve the accuracy of the solutions even when the data is highly corrupted.