Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

Towards Seismic Inverse Problems Using Deep Learning

Jared T. Bryan, Alexander N. Breuer, & Yifeng Cui

Published August 15, 2018, SCEC Contribution #8768, 2018 SCEC Annual Meeting Poster #093

Understanding the behavior and state of fault systems is necessary to make thorough seismic hazard assessments, as well as to realistically model the evolution of earthquake ruptures. Estimation of the evolution of earthquake ruptures is a challenging inverse problem that traditionally requires heavy input of domain experts. We target this problem using deep convolutional neural networks (CNNs) and recurrent neural networks (RNNs), methods whereby the time consuming process of extracting rupture characteristics using traditional inverse methods can be exchanged for the one-time computational cost of training a deep neural network. In order to train a network for the inverse problem, we exploit the well-understood forward problem to generate synthetic seismic data in a realistic two-dimensional domain. The network then learns the mapping from a set of unmigrated seismic records to a description of the rupture front in space and time.

We phrase the inversion first as an image feature extraction, allowing us to use CNNs to map the unmigrated seismic traces to a spatially-discretized fault. In a second experiment, we phrase the inversion as a sequence mapping from unmigrated seismic traces to rupture start and end times for each subfault on a spatially-discretized fault. To train our networks, we took advantage of developments in forward seismic simulations by including a realistic velocity model and complex topography, and exploiting inter-simulation parallelism of the forward solvers. We then utilized parallel, high-throughput GPU computations to train our network.

Our method’s computational costs are accrued entirely up front, after which the model can be applied to unknown data with very little computational power. For our subfault rupture localization and rupture front velocity regression problems, we tested two main network architectures: a deep CNN based on Inception-ResNet-V2 (IRNV2), and a simpler 9-layer CNN. For all tests, we iterate 50 times over 50,000 unique rupture patterns, 70% of which are used for training. IRNV2 performs best on both problems, albeit by a slim margin and at much higher computational cost than the 9-layer CNN. For our subfault rupture initiation and cessation problem, our RNN is able to combine the two problems addressed with CNNs with only slightly decreased accuracy. Further, the application of all three networks to a single dataset provides several independent estimations of the behavior of the earthquake rupture.

Key Words
deep learning, source inversion, rupture inversion, HPC

Bryan, J. T., Breuer, A. N., & Cui, Y. (2018, 08). Towards Seismic Inverse Problems Using Deep Learning. Poster Presentation at 2018 SCEC Annual Meeting.

Related Projects & Working Groups