2025 ASEE Annual Conference & Exposition

Earthquake prediction is an area of interest to researchers around the world as well as
anyone who has experienced a major earthquake. Major earthquakes often cause loss of lives and
property, as well as injuries and destruction. Large investments of time and money are required to
build communities back to near where they were before disasters such as earthquakes. For decades,
scientists have considered different methods for earthquake prediction. Machine learning (ML)
applications have been used in seismology for at least a decade but ML applications in seismology
have increasingly grown during the past few years. In this study, deep learning models will be
applied to three different earthquake datasets, with the goal of predicting earthquake magnitude.
A specific type of recurrent neural network, Long Short-Term Memory (LSTM), with memory
cells that allow for utilizing information form recent past steps, will be applied to earthquake
datasets. These datasets vary in size and are in the form of time-series where earthquake
magnitudes, in Richter scale, are recorded across the time axis. Dataset II is the largest dataset,
containing 50 years of seismic data from 1973/01/02 to 2023/12/31, in a large region that covers
the state of California with a minimum longitude of -133, maximum longitude of -107, minimum
latitude of 24 and maximum latitude of 50. Dataset I is of medium size, covering 3 years of seismic
data from 1970/01/02 to 1973/01/02, in the same region. Dataset III is the smallest dataset which
contains seismic data for 30 days from 2024/05/05 to 2024/06/04. Different sizes of datasets have
been used to study the effect of different timescales. LSTM architectures will be proposed and
tested on three different datasets that are acquired from the United States Geological Survey
Website and their performance will be evaluated and compared. Since earthquakes are natural
phenomena that happen at arbitrary points in time, these time-series are irregular time-series,
meaning the time intervals in between consecutive observations vary in size. To address the
irregularity of the time-series, interpolation will be applied to datasets. It is observed that
interpolation considerably improves the model performance.