Research ArticleGEOLOGY

Machine learning reveals cyclic changes in seismic source spectra in Geysers geothermal field

See allHide authors and affiliations

Science Advances  23 May 2018:
Vol. 4, no. 5, eaao2929
DOI: 10.1126/sciadv.aao2929
  • Fig. 1 Geothermal mechanics.

    (A) General geothermal energy production: Cold water percolates through fracture networks in hot rock, extracts heat, and returns to the surface, where that heat (in vapor) is used to drive a turbine to generate electricity. (B) Qualitative mechanism map of fracture processes that cause earthquakes. Note that the stresses associated with the bottom two processes can cause shear fracture/faulting as well.

  • Fig. 2 Machine learning methods.

    (A) Flowchart of the ML approach, from waveform to fingerprint. NMF and HMM methods both reduce dimensionality and remove features common to all signals. (B) Example of the NMF decomposition of a spectrogram F into the product of matrices, dictionary U, and the activation coefficient matrix diag(a)Vi (notation used in Materials and Methods). (C) In the HMM, each NMF activation matrix can be described as the product of a learned emissions matrix B and a state sequence S. To get to the fingerprint, the algorithm counts every time (t1, t2,…) a certain state follows a previous state; brighter spots mean that one state follows another more frequently. (D) Euclidean (-like) distances are calculated between each fingerprint pair, and then K-means produces J clusters.

  • Fig. 3 Results.

    (A) Map of the Geysers geothermal field region. EGS, Enhanced Geothermal Systems. (B) Map view of earthquake epicenters scaled by magnitude and colored by cluster ID (C1 to C4). (C) Cluster gram shows clear clustering of increased earthquake occurrence in time. Black curve indicates average monthly mass of injected fluid over 71 injection wells at the Geysers geothermal field. Top panel shows total number of earthquakes, which correlates with injection rate (21). (D) Monthly stacked averages over the 3 years in (C) of injection rate (top) and earthquake occurrences in each cluster (bottom), using the kernel density estimator (KDE) of each cluster and averaging the monthly values. The peaks indicate associations of clusters with injection rate (C1, maximum; C4, minimum). C2 and C3 are discussed in the text.

Supplementary Materials

  • Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/4/5/eaao2929/DC1

    section S1. Supplemental text: Algorithms

    section S2. Supplemental text: Further discussion of results

    fig. S1. Distribution of earthquake depths by cluster.

    fig. S2. Energy loss due to attenuation, as a function of distance across the area of the Geysers reservoir, parameterized by Qp and ω or f.

    fig. S3. Distribution of earthquake magnitudes by cluster.

    fig. S4. Kernel density estimators of the histogram values for each cluster.

    fig. S5. HMM with 49 states, station SQK.

    fig. S6. K-means objective function versus J, the number of clusters chosen.

    fig. S7. HMM with 15 states, station SQK.

    fig. S8. HMM with 49 states, station STY.

    fig. S9. Clustering with signals from both stations SQK and STY combined (49 HMM states, 5 clusters).

    fig. S10. Three most characteristic waveforms and associated spectrograms and fingerprints(from top to bottom) from the two clusters associated with maximum (C1) and minimum (C4)fluid injection rates for station SQK.

    fig. S11. Three most characteristic waveforms and associated spectrograms and fingerprints(from top to bottom) from the two clusters associated with maximum (C1) and minimum (C4)fluid injection rates for station STY.

    table S1. List of attached audified sounds and their characteristics.

    audio S1 to S24. Audified seismic data from the three most characteristic earthquakes in each of clusters C1 to C4 (table S1).

    movie S1. Animation of the Geysers earthquake catalog, without and with clustering information.

  • Supplementary Materials

    This PDF file includes:

    • section S1. Supplemental text: Algorithms
    • section S2. Supplemental text: Further discussion of results
    • fig. S1. Distribution of earthquake depths by cluster.
    • fig. S2. Energy loss due to attenuation, as a function of distance across the area of
      the Geysers reservoir, parameterized by Qp and ω or f.
    • fig. S3. Distribution of earthquake magnitudes by cluster.
    • fig. S4. Kernel density estimators of the histogram values for each cluster.
    • fig. S5. HMM with 49 states, station SQK.
    • fig. S6. K-means objective function versus J, the number of clusters chosen.
    • fig. S7. HMM with 15 states, station SQK.
    • fig. S8. HMM with 49 states, station STY.
    • fig. S9. Clustering with signals from both stations SQK and STY combined (49 HMM states,5 clusters).
    • fig. S10. Three most characteristic waveforms and associated spectrograms and fingerprints (from top to bottom) from the two clusters associated with maximum (C1) and minimum (C4)fluid injection rates for station SQK.
    • fig. S11. Three most characteristic waveforms and associated spectrograms and fingerprints (from top to bottom) from the two clusters associated with maximum (C1) and minimum (C4)fluid injection rates for station STY.
    • table S1. List of attached audified sounds and their characteristics.

    Download PDF

    Other Supplementary Material for this manuscript includes the following:

    • audio S1 to S24 (.wav format). Audified seismic data from the three most characteristic earthquakes in each of clusters C1 to C4 (table S1).
    • movie S1 (.mp4 format). Animation of the Geysers earthquake catalog, without and with clustering information.

    Files in this Data Supplement:

Stay Connected to Science Advances

Navigate This Article