- Lin, J. T., Melgar, D., Thomas, A. M., & Searcy, J. (2021). Early warning for great earthquakes from characterization of crustal deformation patterns with deep learning. Journal of Geophysical Research: Solid Earth, e2021JB022703.
- 1. Installation
- 2. Generate rupture and seismic waveforms data
- 3. Run M-LARGE training
- 4. Test M-LARGE model
cd Your_Local_Path
git clone https://github.com/jiunting/MLARGE.gitGo to your environval variable file (.base_profile or .bashrc)
vi ~/.bashrc or
vi ~/.bash_profile and add the following line in the file
#set MLARGE
export PYTHONPATH=$PYTHONPATH:YOUR_PATH_MARGE/MLARGE/src-
Melgar, D., LeVeque, R. J., Dreger, D. S., & Allen, R. M. (2016). Kinematic rupture scenarios and synthetic displacement data: An example application to the Cascadia subduction zone. Journal of Geophysical Research: Solid Earth, 121(9), 6658-6674.
-
Melgar, D., Crowell, B. W., Melbourne, T. I., Szeliga, W., Santillan, M., & Scrivner, C. (2020). Noise characteristics of operational real‐time high‐rate GNSS positions in a large aperture network. Journal of Geophysical Research: Solid Earth, 125(7), e2019JB019197.
You can make your own rupture scenarios and waveforms. The Mudpy source code can be downloaded HERE
Or download the example data directly from
-home_directory
-project_name
-output
-ruptures
-waveforms
*home_directory is the absolute path before the
Chile(e.g. /Users/timlin/Test_MLARGE/)
*project_name isChile
*output, ruptures and waveforms should be exactly the nameoutput; ruptures; waveforms
mkdir Project_Name
cd Project_Name
cp YOUR_PATH_MARGE/example/control.py . #copy example file to your project directoryhome = PATH_To_Project_Name #without the project name
project_name = Project_Name_For_Data #for data, not the Project_Name above
run_name = Prepend_Name>>print(home+project_name+'/output/ruptures/')
>>print(home+project_name+'/output/waveforms/') These should point you to the right directory to .rupt/.log and waveforms directory.
The rupt files are named in run_name.xxxxxx.rupt.
| Variable Name | Meaning |
|---|---|
| save_data_from_FQ | True/False- Write E/N/Z.npy data and STFs from the above waveforms/ruptures directories |
| gen_list | True/False- Generate data path for later training |
| gen_EQinfo | True/False- Read above rupture files and generate rupture information file |
| tcs_samples | Array of the time for ENZ and STF |
| outdir_X | Output directory name for E/N/Z.npy |
| outdir_y | Output directory name for STF |
| out_list | Prepended output name for E/N/Z/STF list file |
| out_EQinfo | Output name for rupture information |
| GFlist | Path of GFlist that used to generate rupture scenarios |
| Sta_ordering | Path of station ordering file (keep always the same order while in training/prediction) |
train_params={
'Neurons':[256,256,128,128,64,32,8],
'epochs':50000,
'Drops':[0.2,0.2],
'BS':128, #Batch Size for training
'BS_valid':1024, #validation batch size
'BS_test':8192, #batch size for testing
'scales':[0,1], #(x-scaels[0])/scales[1] #Not scale here, but scale in the function by log10(X)
'Testnum':'00', #Note use string
'FlatY':False, #using flat labeling?
'NoiseP':0.0, #possibility of noise event
'Noise_level':[1,10,20,30,40,50,60,70,80,90], #GNSS noise level
'rm_stans':[0,115], #remove number of stations from 0~115
'Min_stan_dist':[4,3], #minumum 4 stations within 3 degree
'Loss_func':'mse', #can be default loss function string or self defined loss
}The model architecture is fixed and can only be modified inside the function
mlarge_model.py
#In control.py file
train_model=Truethen,
python control.pyThe trainning model should be saved in the "Testnum" directory
#In control.py file
test_model=Trueand set the inputs
| Variable Name | Meaning |
|---|---|
| Model_path | Path of your model. Load example model by setting Model_path='Lin2020' |
| X | Scaled/unscaled features in dimension of [samples,timesteps,Nfeatures] |
| y | Labels in dimension of [samples,timesteps,1] |
| scale_X | A scaling function f, so that f(X) is ready for training, if X is already scaled, pass a dummy function. |
| back_scale_X | A backward scaling function bf, so that bf(X') is X |
| scale_y | Same as X but scaling function for label |
| back_scale_y | Same as X but scaling function for label |
Finally, test the model by:
>>import mlarge.scaling as scale
>>f = lambda a:a
>>M = mlarge_model.Model(Model_path,X,y,f,scale.back_scale_X,f,scale.back_scale_y) #load model as M
>>M.predict() #make prediction
>>print(M.predictions) # predicted Mw time series
>>print(M.real) #the real Mw
#calculate model accuracy with +-0.3 threshold
>>M.accuracy(0.3)
>>print('Mean model accuracy is {:.2f}%'.format(M.sav_acc.mean())) #model accuracy Mean model accuracy is 97.38%
Note that the accuracy is low at the begining (i.e. 30 sec) and then increases when more data available.
Print all the model accuracy as a function of time.
>>print(M.sav_acc) #plot the accuracy as a function of time
>>M.plot_acc(T=np.arange(102)*5+5,show=True)