On the Use of Neural Networks for Full Waveform Inversion [Software]

Published: 31 January 2023| Version 1 | DOI: 10.17632/7kps2hnj6g.1
Leon Herrmann,
Tim Bürchner,
Felix Dietrich,
Stefan Kollmannsberger


This is the representative code used in the following publication: Herrmann, L., Bürchner, T., Dietrich, F., Kollmannsberger, S., On the Use Neural Networks for Full Waveform Inversion, submitted to Computer Methods in Applied Mechanics and Engineering, February 2023 The driver files are specifically associated with the following sections in the publication: - 4.1.1. Physics-Informed Neural Networks: PINN2D.py - 4.1.2. Physics-Informed Neural Networks with non-trainable Forward Operator: AutomaticDifferentiation2D.py - 4.1.3. Adjoint Method: Adjoint2D.py - 4.1.4. Hybrid Method: Hybrid2D.py - 4.2. Three-Dimensional Case: Adjoint3D.py - 4.2. Three-Dimensional Case: Hybrid3D.py Remark: Despite the manual seed in PyTorch, results are not reproducible across machines. Different initializations are expected in the model input and the initial neural network parameters. Although slight differences in the inversion are expected, the observed tendencies will prevail. To reproduce the results as in the paper, the code can be run on google collab and should lead to exactly the same results.



Technische Universitat Munchen


Inverse Problem, Wave Equation, Inversion, Deep Learning, Neural Network