1 of 45

Applications of Deep Learning in Seismology

Congcong Yuan

10/24/2019

2 of 45

(Kong et al., 2018)

3 of 45

(Kong et al., 2018)

1

2

3

4

5

Data collection

Data preprocessing

Model training

Model evaluation

Production

4 of 45

Detection

Earthquake Seismology (78)

Exploration Seismology (116)

Phase Picking

Location

Early Warning

Prediction

Forward Modeling

Tomography

Others

Denoising

Interpolation

First-break Picking

Tomography

Geologic Interpretation

Others

Microseismic monitoring

Journal publication status (DL in seismology until 2019.6)

5 of 45

Earthquake Seismology

Exploration Seismology

Classification

Regression

  • First-breaking picking
  • Geological interpretation
  • Microseismic monitoring
  • Others
  • Denoising
  • Interpolation
  • Tomography
  • Others

​

  • Detection
  • Phase picking
  • Location
  • Early warning
  • Others

​

​

​

  • Location
  • Focal mechanism
  • Prediction
  • Forward modeling
  • Tomography
  • Others

​

​

CNN/GAN/RF/RNN/…

FCN/DNN/RF/RNN/Autoencoder/…

6 of 45

Signal denoising

Earthquake location

Early warning & prediction

Forward modeling

Subsurface tomography

Seismic phase

detection & picking

Geological interpretation

Main DL applications in seismology

7 of 45

01

Detection & picking

Seismic phase

(Zoss et al., 2018)

8 of 45

01

Detection & picking

Seismic phase

(Zoss et al., 2018)

9 of 45

01

Detection & picking

Seismic phase

(Zhu and Beroza, 2018)

10 of 45

01

Detection & picking

Seismic phase

(Yuan et al., 2018)

11 of 45

01

Detection & picking

Seismic phase

12 of 45

01

Detection & picking

Seismic phase

PcP phase

PKiKP phase

13 of 45

01

Detection & picking

Seismic phase

14 of 45

01

Detection & picking

Seismic phase

Short summary:

​

  1. CNN or RNN-based networks are strong;
  2. It may be promising to replace conventional processes, but need to QC strictly.

Challenges:

​

  1. Stability
  2. Accuracy & QC
  3. How to address all potential phases and deal with complicated cases?
  4. Real applications in seismogram cataloging

15 of 45

(Zhu et al., 2018)

02

Denoising

Seismic signal

16 of 45

02

Denoising

Seismic signal

Denosing convolutional neural network architecture

17 of 45

02

Denoising

Seismic signal

Testing samples

with 50% Guassian noise

18 of 45

02

Denoising

Seismic signal

(Ma et al., 2019)

19 of 45

02

Denoising

Seismic signal

(Gao and Zhang, 2019)

Iteratively data denoising and interpolating

20 of 45

02

Denoising

Seismic signal

(Picetti et al., 2019)

21 of 45

02

Denoising

Seismic signal

Short summary:

​

  1. Deep-learning algorithm may learn the features of target noises and then effectively remove noise from noisy data;
  2. CNN-based algorithms are mostly utilized and GAN produces high-resolution images.

Challenges:

​

  1. QC output and ensure output is high-fidelity;
  2. How to handle various noises in real cases.

22 of 45

03

Location

Passive seismic

(Perol et al., 2018)

(Perol et al., 2018)

23 of 45

03

Location

Passive seismic

(Perol et al., 2018)

(Zhang and Yuan, 2018)

24 of 45

03

Location

Passive seismic

(Perol et al., 2018)

(Zhang and Yuan, 2018)

25 of 45

03

Location

Passive seismic

(Perol et al., 2018)

(Zhang et al., 2019)

26 of 45

03

Location

Passive seismic

(Perol et al., 2018)

(Zhang et al., 2019)

Small earthquake

27 of 45

03

Location

Passive seismic

(Perol et al., 2018)

(Kriegrowski et al., 2018)

28 of 45

03

Location

Passive seismic

(Perol et al., 2018)

Short summary:

​

1. Locating earthquake is possible using CNN-based algorithm;

2. Location accuracy is limited to training data and label accuracy;

3. Deep learning location takes advantage of waveforms rather than traveltime, which may be applied for future earthquake early warning.

29 of 45

04

Early warning & prediction

Earthquake

(Li et al., 2018)

30 of 45

04

Early warning & prediction

Earthquake

(Meier et al., 2019)

31 of 45

04

Early warning & prediction

Earthquake

Whether earthquake larger than magnitude 6 will happen or not (Label 1 – Yes; Label 0 – No)

(Huang et al., 2018)

32 of 45

04

Early warning & prediction

Earthquake

(Rouet-Leduc and Humphreys, 2018)

33 of 45

04

Early warning & prediction

Earthquake

Short summary:

​

  1. Deep learning algorithms can identify short-time (3s) signals with high accuracy;
  2. No promising methods can predict earthquake magnitude with high accuracy, even under magnitude 6.5;
  3. Challenging and opportunity!

34 of 45

05

Modeling & Tomography

Seismic

(Moseley et al., 2018)

35 of 45

05

Modeling & Tomography

Seismic

(Araya-polo et al., 2018)

36 of 45

05

Modeling & Tomography

Seismic

(Yang and Ma, 2019)

37 of 45

05

Modeling & Tomography

Seismic

(Yuan et al., GJI, in review)

38 of 45

05

Modeling & Tomography

Seismic

(Yuan et al., GJI, in review)

39 of 45

05

Modeling & Tomography

Seismic

Short summary:

​

  1. Deep learning modeling may simulate some low-frequency signals in large scale;
  2. Deep learning tomography still depends on dataset and forward solver;
  3. Real applications are still challenging to take place of conventional algorithms.

40 of 45

06

Interpretation

Seismic

(Provided by Dr. Xinming Wu)

41 of 45

06

Interpretation

Seismic

(Provided by Dr. Xinming Wu)

42 of 45

06

Interpretation

Seismic

(Provided by Dr. Xinming Wu)

43 of 45

06

Interpretation

Seismic

(Provided by Dr. Xinming Wu)

44 of 45

06

Interpretation

Seismic

(Provided by Dr. Xinming Wu)

45 of 45

06

Interpretation

Seismic

(Provided by Dr. Xinming Wu)

Short summary:

​

1. CNN-based algorithms are very successful in the geological interpretation;

2. Challenges are: 1. how to handle with complex/uncommon structures; 2. how to build quantitively geological evolutions;