Improving the Earthquake Nowcasting Code
Improving the Earthquake Nowcasting Code #
Comments by Geoffrey Fox, 15 March 2023
The Earthquake forecasting code has been improved following the recent studies of
-
"Does the Catalog of California Earthquakes, With Aftershocks Included, Contain Information About Future Large Earthquakes?" John B. Rundle, Andrea Donnellan, Geoffrey Fox, Lisa Grant Ludwig, James Crutchfield, 10 February 2023 https://doi.org/10.1029/2022EA002521
-
"Optimizing Earthquake Nowcasting With Machine Learning: The Role of Strain Hardening in the Earthquake Cycle," John B. Rundle, Joe Yazbeck, Andrea Donnellan, Geoffrey Fox, Lisa Grant Ludwig, Michael Heflin, James Crutchfield, 17 October 2022 https://doi.org/10.1029/2022EA002343
These two papers identify that the occurrence rate of medium earthquakes
(m > 3.29) represented by smoothing in a time series of their number
reveals the hidden variables controlling large earthquakes with
magnitude >= 6.75. In particular, the rate of these medium earthquakes
decreases before a large quake; due to aftershocks, their number peaks
after a large earthquake.
We added this observable with 9 different smoothing methods to the existing earthquake nowcasting code with the simplest LSTM model. This gave for the earthquake activity in the next 4 years.
-
Normalized Nash-Sutcliffe Efficiency NNSE with 9 "Physics Suggested" Data Training 0.948 Validation 0.866
-
Normalized Nash-Sutcliffe Efficiency NNSE with Original code as in MLCommons Training 0.928 Validation 0.796
Which is a significant improvement. We used the Morris method to find which physics observable was most significant and ran with just this getting slightly better.
- Normalized Nash-Sutcliffe Efficiency NNSE with Best "Physics Suggested" Data Training 0.956 Validation 0.866
The time-dependent four-year predictions for the last fit is
While the original code gives
