Kinematics of Mw6.2 24/August/2016, Amatrice earthquake inferred from the novel fuzzy inversion method
Abstract
The main challenge of kinematic finite fault inversions is the lack of enough observation in the wavefield generated by a given fault rupture. Different kinds of data gathered using different technologies can provide reliable ground-motion data at different frequency bands. One interesting idea to overcome ill-poseness of finite fault inversion is to balance the number of model parameters, and therefore the degree of model flexibility, with the number of available information. The well-known fuzzy function approximation technique, used in machine-learning methods, can adjusts the degree of model flexibility to the complexity of the target function to be approximated. Using this method, we can adjust the size of fuzzy sub-faults considering the observed data.
In this presentation, we will apply the novel fuzzy inversion method (Kheirdast et al, 2019) to infer the rupture kinematics of the Mw6.2 24/August/2016 Amatrice, central Italy. Different data sets, namely static and high-rate GNSS data and strong-motion accelerograms, are used in the learning procedure in order to obtain the parameters of the fuzzy system that approximates the source function. We show that, at low frequencies, a few fuzzy basis functions can perfectly capture the source kinematics. However, at high frequencies (~0.5Hz), the slip becomes highly heterogeneous and more fuzzy basis functions are required. By separating the concepts of parameterization (using the fuzzy basis functions) and numerical accuracy (by the number of points to evaluate the output of fuzzy approximator), we reduce the size of the model-space and increase the precision of modeling. This procedure reveals the best trade-off to filter out the uncertainty (including noise) from the data. In this study, in order to evaluate the solution robustness at high-frequencies (0.06<f<0.5Hz), the observational dataset is divided into the two subsets, namely training and validation. We train the fuzzy system using accelerograms and validate it using high-rate GNSS data. Finally, we compare the similarities of the slip solution obtained using the novel fuzzy approach and those obtained by previously published kinematic/dynamic inversions.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMS053.0007K
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1914 Data mining;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS