error/bug in arrival times

[saeedsltm]

Dear @smousavi05

It seems there is something wrong in detected arrival times (or at least when writing them to a file) as you can see it here in a small portion of Y2000.phs file:

2018 811 65853.0028 31.92 51 17.40 5.000.00
AHRM BI  HHZ IP 02018 811 65846.20        0.00   0
ABTL BP  HHZ IP 02018 811 65852.00        0.00   0
DLRM BP  HHZ IP 12018 811 65852.00        0.00   0
ABAS BP  HHZ IP 12018 811 65853.00        0.00   0
                                                                  200043
2018 811 7 351.5228 31.92 51 17.40 5.000.00
ABTL BP  HHZ IP 12018 811 7 340.00        0.00   0
KHIA BP  HHZ IP 12018 811 7 340.00        0.00   0
NOKA BP  HHZ IP 12018 811 7 342.00        0.00   0
ABAS BP  HHZ IP 02018 811 7 351.00        0.00   0
SHIF BP  HHE     2018 811 7 3 0.00       51.05ES 0
                                                                  200044

1- Decimal seconds are zero for most of P phases (no S phase associated with).
2- Arrival times of some stations for some events, are the same! e.g, look at stations "ABTL" and "DLRM" in the first event or stations "ABTL" and "KHIA" in second one.

[smousavi05]

@saeedsltm HI Saeed. yes there is a minor issue in writing out the decimals, I think I know what causes this and hopefully soon gets some time to fix it.
I think the second issue is related to the first one as for stations with similar station-event distances, the arrival time differences could be as small as a fraction of a second and here losing the decimal points make them look similar.

Could you please send me the X_prediction_results.csv files for these stations?

[smousavi05]

Hi @saeedsltm, I remember you had a dataset on which EqT did not perform well. Blocky Earthquake Transformer or BET(https://github.com/maihao14/BlocklyEQTransformer) might be helpful in that case. BET is a user-interface no-code-based platform that makes it easy to fine-tune EqT for specific data/regions. It also provides a user-friendly interface to create your own EqT model and train it without dealing with coding or knowing in-depth ML. Let me or Hao know if you had any questions.

[saeedsltm]

Dear @smousavi05 Thanks for the suggestion, i will check it out.