3:30 PM - 3:45 PM
[SGD02-06] Fault Geometry Invariance of Crustal Deformation and Streamlined Analysis Using Physics-Informed Deep Learning
In this presentation, we discuss theoretical properties of antiplane dislocations and utilize them to develop streamlined analysis using physics-informed deep learning:
Okazaki T, Hirahara K, Ueda N (2023). Fault geometry invariance for physics-informed crustal deformation learning. (Preprint at: https://doi.org/10.21203/rs.3.rs-3689706/v1)
1. Fault geometry invariance
We state that displacement fields caused by uniform slips on two faults whose fault tips (i.e., dislocation lines) are common are identical up to a constant in a domain enclosed by the two faults. We derive it in a simple and intuitive manner.
2. Dislocation potential
Based on the invariance, we define the dislocation potential as the displacement field caused by the unit slip on the fault connecting a reference point and a given point. We show that the dislocation potential has all information of crustal deformation due to arbitrary fault shapes and slip distributions.
3. Physics-informed deep modeling
We calculate the dislocation potential using a physics-informed neural network (PINN), whose continuous representations in complex structures are suitable for this purpose. We extend PINN crustal deformation modeling (Okazaki et al., 2022) to include the position of dislocation lines, which enables surrogate modeling of crustal deformation due to arbitrary fault slips. We validate its performance in a homogeneous half-space and a heterogeneous structure.
4. Discussion
In this study, we state fault geometry invariance, define dislocation potential, and use it to analyze crustal deformation due to arbitrary fault slips. PINNs are effective owing to its continuous representations in complex structures. A limitation is that this theory holds only for linear antiplane dislocations. This method can be applied to uncertainty quantification and inversion analysis regarding unknown fault geometry.
Okazaki T, Hirahara K, Ueda N (2023). Fault geometry invariance for physics-informed crustal deformation learning. (Preprint at: https://doi.org/10.21203/rs.3.rs-3689706/v1)
1. Fault geometry invariance
We state that displacement fields caused by uniform slips on two faults whose fault tips (i.e., dislocation lines) are common are identical up to a constant in a domain enclosed by the two faults. We derive it in a simple and intuitive manner.
2. Dislocation potential
Based on the invariance, we define the dislocation potential as the displacement field caused by the unit slip on the fault connecting a reference point and a given point. We show that the dislocation potential has all information of crustal deformation due to arbitrary fault shapes and slip distributions.
3. Physics-informed deep modeling
We calculate the dislocation potential using a physics-informed neural network (PINN), whose continuous representations in complex structures are suitable for this purpose. We extend PINN crustal deformation modeling (Okazaki et al., 2022) to include the position of dislocation lines, which enables surrogate modeling of crustal deformation due to arbitrary fault slips. We validate its performance in a homogeneous half-space and a heterogeneous structure.
4. Discussion
In this study, we state fault geometry invariance, define dislocation potential, and use it to analyze crustal deformation due to arbitrary fault slips. PINNs are effective owing to its continuous representations in complex structures. A limitation is that this theory holds only for linear antiplane dislocations. This method can be applied to uncertainty quantification and inversion analysis regarding unknown fault geometry.