EARTHQUAKE PREDICTOR NEURAL NETWORK

(PROJECT ONAMAZU)

| Predictions | Model | Archive | Authorship |

Neural Network Model Information

Initially trained on: 2020-03-17

This is the latest version of the model. Older versions were used in 2018 and 2019.

Dataset:

Source: United States Geological Survey (USGS)
Training Set: 1973-01-01 to 2017-12-31
Test Set: 2018-01-01 to 2018-12-31
Update Set: 2019-01-01 to 2020-03-16
Seed Set: Last 1 Day (Previous 24 Hours)
Temporal Resolution: 1 Day (24 Hours)
Spatial Resolution: 2 Degrees Latitude and 2 Degrees Longitude
Details: Earthquakes were reorganized into tensors, with each slice representing a single day and consisting of a matrix with the rows and columns representing approximate latitudes and longitudes. Each position in the matrix was filled with a magnitude determined by converting all the earthquakes occurring at that approximate location and time from logarithmic scale magnitudes into linear energies, summing them together, and then converting the resulting value back into the logarithmic scale. Some preprocessing is required to convert between different types of magnitude scales as well. This model is designed for the moment magnitudes (Mw). A baseline magnitude of -1.0 is assumed for areas where no event takes place.

Model:

High-Recall Model

Architecture:

Long Short Term Memory (LSTM) Recurrent Neural Network (RNN)
Hidden Layers: 5
Neurons Per Hidden Layer: 512 nodes
Timesteps: 1
Epochs of Training: 1
Loss Function: Asymmetric (Exponential Or Logarithmic)
Library: Keras in Tensorflow 2.1
Notes: Model is fully Stateful and utilizes an Online Learning schedule. Utilizes a dual path architecture that combines dense connectivity with residual skip connections.

Statistics:

Prediction Magnitudes:

All Magnitudes

Either Prediction or Actual Magnitude Is 5.0 Or Greater

(Note: Some statistics cannot be properly assigned because of the lack of true negatives. These numbers are somewhat biased due to the sparsity of predictions above 5.0)