Solving Inverse Problems Using Finite-Element Physics Informed Neural Networks in Presence of Noise
Abstract
This study builds upon a previous investigation of Finite-Element Physics-Informed Neural Networks (FE-PINNs) by performing an analysis of their sensitivity to noise. FE-PINNs were previously shown to be capable of performing a two-dimensional linear elastic full waveform inversion on a soil column. As a further step towards applying this methodology to problems involving real data, FE-PINNs were used to inversely determine the elastic modulus of a single quad element, with varying degrees of noise (0-20%) present in the training data. It was found that, depending on the accuracy of the initial estimate of the element's elastic modulus, FE-PINN can successfully solve the inverse problem with up to 20% noise in the training data.
Solving Inverse Problems Using Finite-Element Physics Informed Neural Networks in Presence of Noise
This study builds upon a previous investigation of Finite-Element Physics-Informed Neural Networks (FE-PINNs) by performing an analysis of their sensitivity to noise. FE-PINNs were previously shown to be capable of performing a two-dimensional linear elastic full waveform inversion on a soil column. As a further step towards applying this methodology to problems involving real data, FE-PINNs were used to inversely determine the elastic modulus of a single quad element, with varying degrees of noise (0-20%) present in the training data. It was found that, depending on the accuracy of the initial estimate of the element's elastic modulus, FE-PINN can successfully solve the inverse problem with up to 20% noise in the training data.