Research
Earthquake Source Processes
Finding patterns in earthquake ruptures is important because they may inform our understanding of why and how earthquakes happen. This is a difficult problem; observations of earthquakes are usually surficial and sparse and are always convolved with path effects from opaque earth structure. Identifying similarities or differences between ruptures requires objective methods that yield distributions of potential source parameters that account for this high level of uncertainty. Our group develops and applies methods of this type.
One focus of ours has been developing a method to compute the covariances of the source distribution in a Bayesian framework. This approach makes few assumptions and yields a distribution of possible rupture behaviors. The manuscript describing this methodology can be found here. This framework allows us to apply the same processing to many events with full posterior distributions for comparability. We are now applying this to a global catalog of events and discovering several important patterns in rupture phenomenology.
Ensemble of map projections of second moment source estimates of spatial extent (left) and directivity (right) for the 2019 M7.1 Ridgecrest earthquake.
Another interest of ours is source time function (STF) deconvolution, for which it has historically been difficult to choose the best empirical Green's functions (EGFs) and evaluate uncertainty. Building on recent advances in deep probabilistic imaging, we are developing tools to objectively perform STF deconvolution in a probabilistic framework. These tools will allow us to compute posterior distributions of STFs and stochastically estimate the optimum EGF.
Volcano seismology
Volcanoes can generate a broad spectrum of seismic signals arising from a diverse suite of processes, including tectonic deformation and magma migration. Volcanic earthquakes are often the best available real-time indicator of subsurface activity before, during, and after an eruption. Our group is working to extend the utility of seismic data in these environments by building high-precision earthquake catalogs, allowing us to image magma system components in unprecedented spatiotemporal resolution. These techniques show promise for enhanced volcano monitoring and provide new insights into key topics in modern volcanology, such as the structures responsible for transporting melt through the deep plumbing system and the extent of magmatic interconnectivity at multi-volcano complexes.
Isometric view of seismicity in the Pāhala Sill Complex, a magma transport structure in the Hawaiian mantle. Individual sills are colored for clarity. Earthquakes are also shown projected into depth sections and map view.
Dynamics of Seismicity
Earthquake sequence dynamics influence our understanding of the physical forces within the Earth. The spatial-temporal patterns of seismicity reveal a rich history of their driving physical mechanisms. There is a conglomerate of patterns suggestive of foreshocks, aftershocks, and other phenomena. A challenging problem is identifying and characterizing the physical processes of these complex earthquake sequences, and whether the source is seismic or aseismic. Recent innovative algorithms in seismic catalog building provide unprecedented detail to study their underlying driving processes and mechanical properties of fault zones.
Our group recently developed an enhanced seismicity catalog for Southern California catalog that led to discovery of widespread multi-year earthquake swarms that exhibit ultra-diffusive patterns consistent with a natural fluid injection process. These highly detailed Cahuilla swarms provide a means to directly image a fault's hydraulic properties. Preliminary work using its extensive spatiotemporal pattern revealed a permeability enhancement in a low-permeability fault zone. Our future research will leverage detailed seismicity catalogs against pioneering techniques to research and model earthquakes and fault processes.
Summary of spatial and temporal behavior of the seismicity during the 2016-2020 Cahuilla swarm.
Ground motion synthesis with generative AI
Reliable ground motion time histories (complete accelerograms) are widely used in engineering design and seismic hazard analysis. However, there are still not enough records of large earthquakes covering the desired range of source, path, and site characteristics. To overcome limitations due to data scarcity, methods for synthesizing ground motion are used to generate ground motions for scenarios of interest. But these traditional stochastic ground motion generation methods have difficulty generating realistic waveforms that span the broadband range of engineering interest (~0.1 – 25 Hz).
Our group has leveraged recent advances in machine learning and high-quality seismic datasets to develop a data-driven ground motion synthesis model for generating 0.1-30Hz, three-component accelerograms conditioned on ground motion parameters. Verification and validation results show that the stochastically synthesized ground motions from our model successfully capture the statistics of ground motion time histories from earthquakes. Our model provides an efficient tool to generate high-fidelity acceleration time histories for engineering applications.
A cartoon summary of our deep generative framework for seismic accelerogram synthesis. A neural operator model (GANO) learns to generate complete three-component earthquake ground motions conditional on a set of input engineering quantities (magnitude, distance, Vs30, faulting type).
Seismic wave propagation and inversion with Neural Operators
Accurate and fast simulation of seismic waves is a major barrier to advancing research in seismology. Conventional numerical solvers are computationally expensive. A new machine learning paradigm called neural operators extends neural networks to learning in infinite dimensional spaces. As universal approximators in function spaces, once trained, neural operators can simulate a full wavefield for arbitrary velocity models and mesh discretization at negligible cost. Our group has pioneered the usage of Neural Operators in seismology and developed models for acoustic and elastic wave equations in 2D and 3D. The models can accelerate wave propagation by orders of magnitude while maintaining good accuracy. The method also enables efficient full-waveform inversion with automatic differentiation.
A Neural Operator architecture for seismic wave propagation. The operator receives a velocity model and source term as input and generates a full 4D seismic wavefield on an arbitrary mesh.