EGU General Assembly 2022
© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

Using K-Means Clustering to Compare Adjoint Waveform Tomography Models of California and Nevada

Claire Doody1, Arthur Rodgers2, Christian Boehm3, Michael Afanasiev3, Lion Krischer3, Andrea Chiang2, and Nathan Simmons2
Claire Doody et al.
  • 1University of California, Berkeley, Berkeley, United States of America (
  • 2Lawrence Livermore National Laboratory, Livermore, United States of America
  • 3Mondaic AG, Zurich, Switzerland

Full waveform inversion models by adjoint methods represent the most detailed seismic tomography models currently available for waveform simulations. However, the influence of starting models on final inversion results is rarely studied due to computational expense. To study this influence, we present three adjoint waveform tomography models of California and Nevada using three different starting models:  the SPiRaL global model (Simmons et al., 2021), the CSEM_NA model (Krischer et al., 2018), and the WUS256 model (Rodgers et al., 2021). Each model uses the same dataset of 103 events between magnitudes 4.5 and 6.5 that occurred from January 1, 2000 to October 31st, 2020. For each event, 175-475 stations record data, creating dense path coverage over California. The model iterations are computed using Salvus. We begin by  running iterations for each starting model at three period bands: 30-100 seconds, 25-100 seconds, and 20-100 seconds. For each period band, we run iterations until the average misfit for all events is no longer reduced; over all period bands, we run more than 55 iterations and see misfit reductions of up to 40% in some period bands. Each model shows velocity anomalies of up to 20%, but the difference in VS values between the models can be significant. Most of these differences seem to correlate with small-scale differences in the starting models. To test whether these differences between the models could affect the interpretation of their results, we utilize k-means clustering to analyze the similarities in large-scale structure in all three models (e.g. Lekic and Romanowicz, 2011). We separate each model into a crustal layer (0-30km depth) and uppermost mantle layer (30-150km), then run a k-means clustering algorithm on absolute Vs wavespeeds and anisotropy [(Vsh/Vsv)^2] separately. We show that regardless of the differences seen on visual inspection, all three models can resolve tectonic-scale structures equally.


This work was supported by LLNL Laboratory Directed Research and Development project 20-ERD-008. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL-ABS-830615.

How to cite: Doody, C., Rodgers, A., Boehm, C., Afanasiev, M., Krischer, L., Chiang, A., and Simmons, N.: Using K-Means Clustering to Compare Adjoint Waveform Tomography Models of California and Nevada, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3257,, 2022.


Display file

Comments on the display

to access the discussion