HARPA: High-Rate Phase Association with Travel Time Neural Fields

About

This is the source code for paper High-Rate Phase Association with Travel Time Neural Fields.

Installation

Simply run:

pip install -q git+https://github.com/DaDaCheng/phase_association.git

It requires torch,obspy, pyproj, pandas, POT and skfmm (only from unknown wave speed model).

Quick start in SeisBench and colab demo

HARPA seamlessly integrates with workflow in SeisBench.

Examples
2019 Ridgcrest earthquake
2014 Chile earthquake
Demo for unknown wave speed and neural fields

Usage

To perform the association, use the following code:

from harpa import association
pick_df_out, catalog_df = association(pick_df, station_df, config)

Input Requirements

pick_df

A DataFrame containing seismic pick information with the following structure:

id	timestamp	prob	type
CX.MNMCX.	2014-04-01 00:08:45.410	0.113388	p
CX.MNMCX.	2014-04-01 00:10:56.840	0.908900	p
CX.MNMCX.	2014-04-01 00:13:01.930	0.795285	p
CX.MNMCX.	2014-04-01 00:13:36.950	0.130214	s

• id: Station id
• timestamp: arrival time of the picks (np.datetime64)
• prob: Optional column indicating the probability of the pick.
• type: Indicates the phase type (e.g., p or s).

station_df

A DataFrame containing seismic station information with the following structure:

id	x (km)	y (km)	z (km)
CX.PB01.	449.359240	7672.990624	-0.900
CX.PB02.	407.073618	7642.202105	-1.015
CX.PB03.	422.289208	7561.616351	-1.460
CX.PB04.	381.654501	7529.786448	-1.520

---

Configurations and Notes

Mandatory Configurations

Config	Description	Example
x(km)	X range for the search space	[0, 100]
y(km)	Y range for the search space	[0, 100]
z(km)	Z range for the search space	[0, 100]
vel	Homogeneous wave speed model	{"P": 7, "S": 4.0}

Optional Configurations

Config	Description	Default
lr	Learning rate	0.1
noise	$\epsilon$ in SGLD	1e-3
lr_decay	Learning rate and noise decay	0.1
epoch_before_decay	Number of epochs before decay	3000
epoch_after_decay	Number of epochs after decay	3000
LSA	Use linear sum assignment in computing loss	True
wasserstein_p	Order of the Wasserstein distance	2
P_phase	Indicates if data contains P-phase	True
S_phase	Indicates if data contains S-phase	False
noisy_pick	Indicates if data contains missing/spurious picks	True
min_peak_pre_event	Minimum number of picks per event	16
min_peak_pre_event_p	Minimum number of P-phase picks per event	0
min_peak_pre_event_s	Minimum number of S-phase picks per event	0
max_time_residual	Maximum allowable time residual for events	2
denoise_rate	Filtering based on nearest station ratio	0
DBSCAN	Use DBSCAN to segment data into different slices	True
remove_overlap_events	Remove repeated events if slices overlap	False
neural_field	True for unknown wave speed model, False otherwise	False
wave_speed_model_hidden_dim	Hidden dimension for wave speed model if neural_field=True	-
optimize_wave_speed	Optimize wave speed model	False
optimize_wave_speed_after_decay	Optimize wave speed after learning rate decay	True
second_adjust	Adjust locations after training	True
time_before	Time difference between the search start and first pick	0.5 * maximum_search_distance / P_wave_speed

---

Notes

1. Multiprocessing and Seisbench Compatibility:

   • The multiprocessing module might conflict with Seisbench's multithreading. If you use Seisbench first, wrap your association call as follows:

   from concurrent.futures import ThreadPoolExecutor
   with ThreadPoolExecutor(max_workers=100) as executor:
       future = executor.submit(association, pick_df, station_df, config, verbose)
   pick_df_out, catalog_df = future.result()

   Otherwise, you can directly use:

   pick_df_out, catalog_df = association(pick_df, station_df, config)

2. Training Epochs:

   • Fixed wave speed models require fewer training epochs. Unknown wave speed models may need more epochs.
   • Usually, for a fast association, we set epoch_before_decay=epoch_after_decay=1000. For a detailed association (for more events) we set epoch_before_decay=epoch_after_decay=10000. For unknown wave speed models, we set epoch_before_decay=epoch_after_decay=10000. A better early stop strategy will be implemented in the next version.

3. Using Unknown Speed Models:

   • For unknown speed models, include a model_traveltime parameter:

   pick_df_out, catalog_df = association(pick_df, station_df, config, model_traveltime=model_traveltime)

   • Here, model_traveltime is a model with the source location as input and travel times to each station as output. See the example for details.

4. Verbose Settings:

   • verbose > 10: print picks and stations.
   • verbose > 8: print training details for each epoch.
   • verbose > 6: print training settings, including CPUs.
   • verbose > 4: print configuration details.
   • verbose > 2: print other details.
   • verbose = 0: silent mode.

---