Introduction

  • Seismic hazards cause numerous deaths and huge financial losses.
  • Earthquake forecasting is known to be extremely difficult.
  • Often seen as elusive and perhaps impossible.beroza2021machine
  • The mechanisms that lead to earthquakes are not sufficiently well known.stein2003introduction
  • Mantle convection plays a major role in the movement of tectonic plates.
  • Convection causes elastic energy to build up on the faults.
  • Elastic-rebound theory:
    • Relative plate movement occurs until it is locked.
    • Elastic strain energy then builds up on the fault.
    • Eventually a fracture occurs, which can result in an earthquake.
Adapted from the original. Original by Xiaohan Song.

Sunspots

  • Sunspots are temporary dark areas on the Sun's surface.
  • They are caused by intense, concentrated magnetic fields that inhibit convection.
  • Indicators of Solar Activity
    • High sunspot numbers are linked to more frequent solar flares and Coronal Mass Ejections.
    • Both release charged particles that disrupt our magnetosphere.
    • High sunspot activity raises the Sun's Total Solar Irradiance, which can influence Earth's climate and surface temperatures.hempelmann2012correlation
Source: Solar and Heliospheric Observatory, NASA

Edit Distances

  • Let \(P_1\) and \(P_2\) be two marked point processes: $$ P_1 = \{ (t_i, \vec{u}_i) \mid 1 \le i \le N_1 \} \quad\quad P_2 = \{ (s_j, \vec{v}_j) \mid 1 \le j \le N_2 \} $$
  • The idea is to transform \(P_1\) into \(P_2\) using primitive operations, each incurring a cost.
  • The primitive operations are:
    • Insertion: insert point \((s_j, \vec{v}_j)\) from \(P_2\) into \(P_1\), paying a cost of 1.
    • Deletion: remove point \((t_i, \vec{u}_i)\) from \(P_1\), paying a cost of 1.
    • Shifting: replace point \((t_i, \vec{u}_i)\) in \(P_1\) by the point \((s_j, \vec{v}_j)\) from \(P_2\), paying a cost of: $$ C\big((t_i, \vec{u}_i), (s_j, \vec{v}_j)\big) = \lambda_0 |t_i - s_j| + \sum_{k=1}^{d} \lambda_k \big| \vec{u}_i(k) - \vec{v}_j(k) \big| $$
  • The \(\lambda_k\) are user-defined normalizing constants intended to ensure that the different marks have comparable orders of magnitude.
  • The edit distance is defined as the lowest possible cost necessary to transform \(P_1\) into \(P_2\).
  • Shift cost:
  • $$ \lambda_0 |t_i - s_j| + \sum_{k=1}^{d} \lambda_k \big| \vec{u}_i(k) - \vec{v}_j(k) \big| $$
  • i.e. the effort to move the circle to its final position
  • Deletion cost: 1 (per node)
  • Insertion cost: 1 (per node)
  • Number of operations:
    • Shifting: 7
    • Deletion: 2
    • Insertion: 2

Radial Basis Functions

  • We use these edit distances with Radial Basis Functions (RBFs) for performing the forecasting.
  • Simply put, it finds a functional relation of the form: $$ \hat{y}(\vec{x}, \vec{w}) = w_0 + \sum_{i = 1}^{k} w_i \, \psi\big(d_H(\vec{x}, \vec{c}_i)\big) $$
  • Where:
    • $\vec{x}$: the point process (e.g., the set of earthquakes in a week) for which we want to make a prediction.
    • $\vec{c}_i$: point processes taken from the training set, which will be compared with $\vec{x}$.
      • We take $100$ randomly chosen instances $\vec{c}_i$ from the training set.
    • $w_i$: adjustable parameters found by least squares.
    • $\psi$: taken to be the Gaussian kernel, $\psi(x) = \exp(-x^{2} / \epsilon)$.
  • We predict the logarithm of the number of earthquakes, $\log(N + 1)$, in a certain day.
  • The predicted values of $\log(N + 1)$ are then correlated with the maximum magnitude by means of Gutenberg-Richter's law.stein2003introduction

Dynamical Systems: A Brief Introduction

  • A dynamical system is simply a system whose state evolves over time according to a fixed rule.
  • Core Components:
    • State: A snapshot of the system at a specific time (e.g., position and velocity of a pendulum).
    • State Space: The set of all possible states the system can be in.
    • Evolution Rule: A mathematical rule ($f$) that describes how the state $x$ changes over time. $x_{t+1} = f(x_t)$
  • Common examples include planetary orbits, weather patterns, and population dynamics.
  • Many useful tools in this field:
    • State-space reconstruction based on time series of observations
    • Prediction by nearest-neighbors
    • Tools to detect coupling between two systems
    • and a lot more

Example: Lorenz System

  • The system of ordinary differential equations:
$$ \begin{aligned} \dot{x} & = \sigma(y-x) \\ \dot{y} & = \rho x - y - xz \\ \dot{z} & = -\beta z + xy \end{aligned} $$

Example: Hénon Map

  • The map function:
$$ \begin{aligned} x_{n+1} &= 1 - ax_n^2 + y_n \\ y_{n+1} &= bx_n \end{aligned} $$
  • Numerous dynamical systems could be involved in the earthquake generating process.
  • One example are the convection flows within the Earth's mantle (Navier-Stokes equations).
  • The figure below shows an exaggerated diagram.

Figure author: Surachit

Coupling Detection Methods

  • One dynamical system can have a strong influence in another (coupling).
  • To measure coupling, we use two methods:
    • Hirata et al. (2016)hirata2016plosone
    • Andrzejak et al. (2011)andrzejak2011characterizing
  • They both rely on the same basic idea:stark1999
    • Measure systems $X$ and $Y$ at times $t_0$ and $t_1$, observing $s_X^{t_0}$, $s_X^{t_1}$, $s_Y^{t_0}$ and $s_Y^{t_1}$
    • $s_Y^{t_0}$ and $s_Y^{t_1}$ are similar $\overset{\text{?}}{\Rightarrow}$ $s_X^{t_0}$ and $s_X^{t_1}$ are also similar
    • Then we say system $X$ drives system $Y$
    • Measure systems $X$ and $Y$ at times $t_0$ and $t_1$, observing $s_X^{t_0}$, $s_X^{t_1}$, $s_Y^{t_0}$ and $s_Y^{t_1}$
    • $s_Y^{t_0}$ and $s_Y^{t_1}$ are similar $\overset{\text{?}}{\Rightarrow}$ $s_X^{t_0}$ and $s_X^{t_1}$ are also similar
    • Then we say system $X$ drives system $Y$

Effects of Lunar (and Solar) Tidal Forces

  • The difficulty of earthquake forecasting is aggravated by the fact that:
    • we cannot yet enumerate all the external factors influencing earthquakes
    • or the degree of importance of each of these factors
  • In the master's dissertation, the connection between Sun activity & earthquakes was analyzed.
  • Here we continue investigating external factors that affect earthquakes
    • We began by analyzing tidal forces from the Moon
    • Solar tidal forces were analyzed, but were not as fruitful
  • The influence of tidal forces has been investigated in the literature ide2016naturegeosciencehao2019evidence
  • However, either:
    • they analyze only correlation, without assessing its applicability in prediction;
    • or they perform predictions for larger time frames (above 15 days)
  • Tidal forces are theorized to influence earthquakes due to its ability to deform the continental crust
  • This could play a significant role in the rate of elastic strain energy accumulation in faults
  • Most importantly, it could be a significant agent that causes the initial crack (nucleation) that develops into an earthquake
  • We simulated the trajectories of the Sun and Moon using the Python library Astropy.
  • We calculated:
    • hourly differential pulls on each region
    • differential pull at the place and time of each earthquake in our catalogs
  • To include sunspot and Moon tidal force data into our prediction framework, we modify our framework as follows:
  • $$ \hat{z}(\vec{x}, \vec{y}^{(1)}, \vec{y}^{(2)}, \vec{w}) = w_0 + \sum_{i = 1}^k \bigg[ w_i \, \psi\big( d_X(\vec{x}, \vec{c}_i) + \kappa_1 d_E(\vec{y}^{(1)}, \vec{y}^{(1)}_{\vec{c}_i}) + \kappa_2 d_E(\vec{y}^{(2)}, \vec{y}^{(2)}_{\vec{c}_i}) \big) \bigg], $$
  • Where:
    • $\vec{y}^{(1)}$: vector containing the 7 sunspot numbers referring to the same week as $\vec{x}$
    • $\vec{y}^{(1)}_{\vec{c}_i}$: same as above, but referring to the same week as $\vec{c}_i$;
    • $\vec{y}^{(2)}$: the $168$-dimensional vector containing the hourly tidal force collected over the 7-day window
    • $d_E()$: the Euclidean distance

Tidal Force Calculation: Differential Pull

  • What is commonly called tidal force is more accurately a differential pull.
  • It arises because the Moon's gravitational pull has varying intensity across the Earth's diameter.
  • The approximate differential pull on an object at the Earth's surface is given by:sawicki1999myths $$ \frac{2 G m_M r_E}{d_M^3} $$
  • A similar effect occurs on both the near and far sides of the Earth.

Causality: Andrzejak et al. Methodandrzejak2011characterizing

  • Found evidence of Moon influencing earthquakes ($Y \rightarrow X$).
  • Coupling measure $|L(Y \rightarrow X)| \gg |L(X \rightarrow Y)|$.
  • Histograms for $L(Y \rightarrow X)$ (left) show irregular, slanted distributions, indicating coupling.
  • Histograms for $L(X \rightarrow Y)$ (right) are nearly uniform, indicating no coupling.
Histogram for New Zealand coupling test Histogram for Balkan region coupling test

Causality: Hirata et al. Methodhirata2016plosone

  • Unidirectional coupling was also identified using this method.
  • Generally, $p$-values for the Moon-to-earthquake ($Y \rightarrow X$) direction were much smaller than for the reverse direction.
  • Exception: The Japan catalog, where the relation was inverted, possibly due to stochastic factors.
  • Overall Conclusion: The collective results from both methods provide evidence for a causal coupling from the Moon to earthquakes.

Correlation: Tidal Force vs. Magnitude

  • We also investigated the direct relationship between differential pull and earthquake magnitude.
  • This analysis focuses on correlation to corroborate the causality findings.
  • We calculated the average differential pull for earthquakes with magnitude $\ge m$, for various values of threshold $m$.
Chart showing average differential pull increases with earthquake magnitude threshold

Correlation Findings

  • Clear Trend: The average differential pull increases with the magnitude threshold across all catalogs.
  • This indicates a tendency for larger-magnitude earthquakes to occur during periods of higher tidal force.
  • Main Finding: A persistent positive correlation exists between lunar tidal force and earthquake magnitudes.
  • We also compared the differential pulls of the 20% lowest magnitude earthquakes against those above various higher thresholds (e.g., >5, >5.5, >6).
  • The earthquakes with larger magnitudes have larger differential pulls.
    • $p < 0.05$ for ~70% of the cases, with a Wilcoxon rank-sum test
Chart showing average differential pull increases with earthquake magnitude threshold

Improving Forecasting Accuracy

  • Goal: Verify if Sun (sunspots) and Moon (differential pulls) data improve forecasting accuracy.
  • Method:
    • Forecast the log-number of earthquakes: $\log(N+1)$.
    • This is then correlated with maximum magnitudes (Gutenberg-Richter law).
Example predictions for New Zealand, showing correlation between predicted log-number of earthquakes and real magnitudes.

Example forecasts for New Zealand

Forecasting Results: Japan

Japan (Nationwide)
Boxplot of correlation for Japan nationwide predictions.
Boxplot of odds-ratio for Japan nationwide predictions.
  • Moon data slightly improves correlation.
  • Best correlation gain (+5.37%) with Moon + Sun data.
  • Odds-ratio only improves (+3.51%) when both are combined.
Touhoku Region
Boxplot of correlation for Touhoku region predictions.
Boxplot of odds-ratio for Touhoku region predictions.
  • Higher baseline accuracy.
  • Significant accuracy gains.
  • Correlation up to +29.62%.
  • Odds-ratio up to +76.63%.

Forecasting Results: New Zealand & Balkan Region

New Zealand
Boxplot of correlation for New Zealand predictions.
Boxplot of odds-ratio for New Zealand predictions.
  • Sunspot data most relevant for correlation (+8.86%).
  • Combined data yielded best odds-ratio (+5.72%).
Balkan Region
Boxplot of correlation for Balkan region predictions.
Boxplot of odds-ratio for West Balkan subregion predictions.
  • Smaller subregion yielded higher correlations.
  • Substantial increases in both metrics from Sun or Moon data.
  • The highest accuracy is achieved when all data types were used simultaneously.

Conclusion: Summary of Findings

  • Presented significant evidence that seismic activity is influenced by tidal forces from the Moon.
  • Moon's Influence:
    • Established a unidirectional causal link (Moon's motion drives earthquakes).
    • Found a slight-to-moderate tendency for larger earthquakes to occur during periods of higher tidal force.
  • Forecasting: Demonstrated that next-day magnitude forecasting accuracy improves when using lunar data, and even more so when combining lunar and sunspot information.

Pinpointing the Sun-Earthquake Mechanism

  • Our previous work established a causal link: solar activity influences earthquakes.junqueira2022solar

  • However, the physical mechanism behind this link is still an open question. Potential pathways include:
    • Interactions with the geomagnetic field
    • Emission of solar particles
    • Tidal forceside2016naturegeoscienceyabe2015tidal

  • Our hypothesis: The thermal energy (heat) transfer from the Sun is a major player in driving changes in the Earth's surface temperature.
    • If it is a major player, then we can say it is being neglected in the literature.

Potential Mechanisms by Which Heat Could Affect Earthquakes

  • Thermal Stress
    • Temperature variations can induce thermal stress in rocks, altering their mechanical properties and contributing to fracture kang2019review.
    • While surface heat penetration is slow, it can affect deeper layers over long timescales.kalogirou2004measurements
  • Hydrological Cycle
    • Atmospheric temperature variations influence the hydrological cycle (evaporation, precipitation).
    • This impacts subsurface water flow and pore pressure within faults, potentially triggering earthquakes.wang2021earthquakes
    • Seasonal changes like snow/ice melting can alter pore pressure and load the Earth's crust.
    • Example: Fluid migration from snowfall may have triggered the M7.6 Noto, Japan earthquake in 2024.wang2024untangling
  • Atmospheric Pressure
    • Changes in atmospheric pressure can induce small-scale clamping and unclamping of faults.liu2009slow
    • A recently identified phenomenon is "stormquakes", where intense storms generate seismic waves in the crust.fan2019stormquakes

  • Thus, several plausible mechanisms have been proposed.
  • However, the supporting evidence for these mechanisms remains largely inconclusive.
  • This shows a clear need for further investigation into this area.

Seasonal Variation of Earthquake Location

  • Analyzed the correlation between earthquake latitude and magnitude in the northern hemisphere, differentiated by season.
  • A statistically significant lower correlation is observed during the winter ($p = 0.002$ via t-test).
  • This indicates that larger earthquakes have a slight tendency to occur at higher latitudes during all seasons except for winter.
Boxplot showing the correlation between earthquake latitude and magnitude across different seasons.
  • This result points to a driving force connected to both latitudinal position and the cycle of meteorological seasons.
  • Solar heat fits this description well, strengthening the hypothesis that it is a significant mechanism for the Sun's influence on earthquakes.

Are Solar Activity & Earthquakes Deterministic or Stochastic?

  • We employed recurrence plots to test whether the time series are deterministic or stochastic.
  • Methodology from Hirata, Y. (2021)hirata2021recurrence was used to count the variety of recurrence triangles.
  • The method indicates that both solar activity and earthquake occurrences are stochastic.
  • We conjecture that a portion of the stochasticity in seismic activity is inherited from the Sun
Illustration of a recurrence plot and recurrence triangles.
Illustration of a recurrence plot and recurrence triangles.
Graph showing super-exponential growth of recurrence triangle varieties for sunspots and earthquakes.
Super-exponential growth of recurrence triangle varieties for sunspots and earthquakes.

Mutual Information Confirms Shared Stochasticity

  • Mutual information was used to quantify the statistical dependency (linear and non-linear) between the two systems.
  • $$ I(X; Y) = \sum_{x \in X} \sum_{y \in Y} p(x, y) \log \left( \frac{p(x, y)}{p(x)p(y)} \right) $$
  • Mutual information is non-zero and significantly larger than that calculated with random noise.
  • This demonstrates that information is shared between the two systems, reinforcing the previous conjecture.
Graph of mutual information between sunspots and earthquakes with a time delay.
Graph of mutual information between earthquakes and a random recurrence plot.

The Sun-Earthquake Time Lag

  • A key aspect of the Sun-earthquake relationship is the time lag between cause and effect, which has been largely overlooked in the literature.
  • Mechanisms for solar influence require significant time to manifest.
    • e.g. Thermal energy from the Sun may take weeks or months to accumulate and impact secondary processes (e.g., ice melting, atmospheric pressure changes).pan2023annualseneviratne2010investigating
  • Research Question: What is the time lag between changes in solar activity and their subsequent effects on Earth's seismicity?
  • Used daily sunspot numbers with various time lags to forecast maximum earthquake magnitudes.
  • The plots show how forecasting accuracy changes when sunspot data is time-lagged.
Prediction accuracy vs. sunspot delay for Japan.
Prediction accuracy vs. sunspot delay for Greece.
Prediction accuracy vs. sunspot delay for New Zealand.
  • Lags that offer statistically significant improvements (Bonferroni corrected $p < 0.05$):
    • Japan: 30, 120, 180, 240, 300 days
    • Greece: 30, 90, 120, 150, 180, 300 days
    • New Zealand: 60, 90, 120, 150, 180, 300, (...) days
  • A delay of 120 days is the first significant lag common to all three regions, possibly related to seasonal changes.
  • Recall that the mutual information for a zero-day lag was substantially lower than the peak values observed at non-zero lags.
  • This further supports the existence of a delayed cause-effect relationship.
Graph of mutual information between sunspots and earthquakes with a time delay.
  • Thus, in summary:
    • A time lag exists between changes in solar activity and their effects on terrestrial seismic activity.
    • A single, universal value for this delay does not exist; it is region-specific.
    • The existence of a range of effective delays is physically plausible due to the multiple mechanisms by which thermal energy can trigger earthquakes (e.g., ice melt).

Forecasting With Surface Temperatures: Japan

  • Data on daily average air temperatures from Tokyo, Stylida, and Wellington can, in most cases, improve the prediction of next-day maximum earthquake magnitudes.
  • The results clearly indicate that incorporating surface temperature data helps to improve prediction accuracy.
  • Combining both sunspot and temperature data achieved the highest correlation among all settings tested for this region.
Japan forecast accuracy for any depth any depth
Japan forecast accuracy for depth < 100km depth < 100km
Japan forecast accuracy for depth < 50km depth < 50km
Japan forecast accuracy for depth < 30km depth < 30km

  • The results strengthen the idea of a causal connection between surface temperatures and earthquake magnitudes.
  • The incremental improvement from using both datasets is modest, indicating informational redundancy.
  • This redundancy suggests that the Sun influences earthquakes via the same thermal pathway by which it drives surface temperatures.

Forecasting Results: Balkans

  • The inclusion of sunspots overall provided better improvements than surface temperatures.
  • The highest correlation for this region was achieved when both sunspot and temperature data were used in conjunction.
  • The biggest improvements were observed when restricting the dataset to shallow earthquakes
Balkans forecast accuracy for any depth any depth
Balkans forecast accuracy for depth < 100km depth < 100km
Balkans forecast accuracy for depth < 50km depth < 50km
Balkans forecast accuracy for depth < 30km depth < 30km

  • The results strengthen the idea of a causal connection between surface temperatures and earthquake magnitudes.
  • The incremental improvement from using both datasets is modest, indicating informational redundancy.
  • This redundancy suggests that the Sun influences earthquakes via the same thermal pathway by which it drives surface temperatures.

Forecasting Results: New Zealand

  • Using temperature data yields a correlation that is not statistically different from the baseline.
  • Using sunspot data results in a significantly better correlation.
  • The importance of temperature data becomes more pronounced as earthquake depths are restricted to be shallower.
New Zealand forecast accuracy for any depth any depth
New Zealand forecast accuracy for depth < 100km depth < 100km
New Zealand forecast accuracy for depth < 50km depth < 50km
New Zealand forecast accuracy for depth < 30km depth < 30km

  • The results strengthen the idea of a causal connection between surface temperatures and earthquake magnitudes.
  • The incremental improvement from using both datasets is modest, indicating informational redundancy.
  • This redundancy suggests that the Sun influences earthquakes via the same thermal pathway by which it drives surface temperatures.

Variations Depending on Earthquake Depth

  • Many natural phenomena that influence earthquakes (ice melt, sea-level changes, storms) primarily exert their effects at shallow depths.
  • These phenomena are often linked to atmospheric temperature and sunspot activity.
  • Idea: If temperature and/or sunspots influence earthquakes, the intensity of this effect should vary with earthquake depth.
Japan
Balkans
New Zealand
Japan forecast accuracy for any depth
Japan forecast accuracy for depth < 100km
Japan forecast accuracy for depth < 50km
Japan forecast accuracy for depth < 30km
Balkans forecast accuracy for any depth
Balkans forecast accuracy for depth < 100km
Balkans forecast accuracy for depth < 50km
Balkans forecast accuracy for depth < 30km
New Zealand forecast accuracy for any depth any depth
New Zealand forecast accuracy for depth < 100km depth < 100km
New Zealand forecast accuracy for depth < 50km depth < 50km
New Zealand forecast accuracy for depth < 30km depth < 30km
Japan forecast accuracy for any depth
Japan forecast accuracy for depth < 100km
Japan forecast accuracy for depth < 50km
Japan forecast accuracy for depth < 30km
Balkans forecast accuracy for any depth
Balkans forecast accuracy for depth < 100km
Balkans forecast accuracy for depth < 50km
Balkans forecast accuracy for depth < 30km
New Zealand forecast accuracy for any depth
New Zealand forecast accuracy for depth < 100km
New Zealand forecast accuracy for depth < 50km
New Zealand forecast accuracy for depth < 30km
A sharp increase in correlation is observed as the depth threshold is decreased.
In absolute value, the greatest improvement in Temp & Sunspots comes when using depth < 30km.
In almost all cases, Temp & Sunspots yields the best prediction accuracy.
When depth < 30km, Temp & Sunspots always yields the best accuracy.

Analyzing the Evidence

  • The idea that atmospheric temperature affects earthquakes is accepted, but is the influence significant or negligible? Our results provide insights.
  • Seasonal Variation:
    • The existence of seasonal variation in earthquake locations points to the importance of seasonal drivers like atmospheric temperatures, ice melt, and storms.
  • Delayed Cause-and-Effect:
    • The time lag between solar activity and seismic response is on the order of 1 to several months.
    • This aligns with long-term effects of surface temperature penetrating the Earth's crust.
    • Surface temperature data improves predictability, strengthening this hypothesis.
    • Informational redundancy between solar and temperature data suggests a common thermal pathway.
  • Shallow Earthquakes:
    • Predictability improves when analysis is restricted to shallow earthquakes.
    • This is consistent with surface-level phenomena (driven by temperature) primarily affecting the upper crust.
  • To sum up:
    • We found solid evidence, from various different kinds of analysis, showing the significance of thermal energy.
    • It is on par, if not greater, than other possibilities.
    • Future work should work on narrowing it down further (hydrological effects, atmospheric pressure, etc).

Ongoing Research: Neural Networks in Earthquake Forecasting

  • We have extensively analyzed the applications of edit distances.
    • Edit distances allow us to extract a very particular kind of information from earthquakes.
  • How can we concilliate our results with what is in the literature (e.g. statistical indicators)?
  • Neural Networks are flexible tools that could help us achieve that.
  • However, a common trend is to progressively increase model size and depth, hoping greater capacity yields better performance.

Author: BrunelloN on WikiMedia

A Problematic Tendency

  • From a Statistical Learning Theory (SLT) perspective, this is an issue.luxburg2011statistical
  • Expanding a network's architecture increases its capacity (e.g., VC dimension), heightening the risk of overfitting.vapnik1998statistical
  • The common response is regularization (dropout, early stopping, etc.).
  • However, this is not a complete solution for seismic forecasting:
    • With limited and noisy data, even well-regularized large models can produce unstable results.
    • Heavy regularization on oversized models can be difficult to interpret.

Author: BrunelloN on WikiMedia

The Core Problem in Statistical Learning Theory (Opt.)

  • The goal of a learner is to find a function \(f\) that minimizes the true risk (generalization error): \[ R(f) = \mathbb{E}_{(x,y)\sim\mathcal{D}}\big[\,\ell(f(x),y)\,\big] \]
  • However, we can only minimize the empirical risk on our training sample \(S\): \[ \hat{R}_n(f) = \frac{1}{n}\sum_{i=1}^n \ell(f(x_i),y_i) \]
  • Central Question of SLT: Under what conditions does a small empirical risk imply a small true risk?
  • The answer depends on the space of admissible functions \(\mathcal{F}\) (the set of all possible models the learner can choose from).
  • The generalization gap is bounded by the complexity, or capacity of \(\mathcal{F}\): \[ \big|R(f) - \hat R_n(f)\big| \le \mathcal{N}(\text{capacity of } \mathcal{F}, n) \]
  • A larger, more complex \(\mathcal{F}\) increases the risk of overfitting.

The Importance of Capacity Control (Opt.)

  • This leads to the classic bias-variance trade-off:
    • If \(\mathcal{F}\) is too small (low capacity): high bias (model can't capture the underlying pattern).
    • If \(\mathcal{F}\) is too large (high capacity): high variance (model fits noise in the training data).
  • The key is to precisely control \(\mathcal{F}\) so it's small enough to prevent overfitting, but large enough to contain the function we're looking for.
  • Two Strategies for Capacity Control:
    1. Directly restrict the space \(\mathcal{F}\): Our primary approach. Design a smaller, dedicated architecture (e.g., fewer parameters, constrained connectivity).
    2. Regularization: Penalize complexity while searching in a larger space (e.g., \(\ell_1, \ell_2\), dropout, early stopping). A complementary, but less direct, technique.
  • Our Argument: Many ML approaches for earthquake forecasting use unnecessarily large \(\mathcal{F}\) and rely only on regularization.

Space of Admissible Functions: $\mathcal{F}$ (Opt.)

𝓕

A Capacity-Aware NN Architecture

  • Motivated by SLT, we designed a custom neural network with deliberately reduced capacity.
  • How? By constraining the first layer to be not fully connected, we reduce the space of admissible functions \(\mathcal{F}\) without sacrificing expressiveness.
  • Instead of starting with a large model and taming it, we design models whose capacity is constrained to reflect our domain knowledge.
  • Core Innovations:
    1. Calculating seismicity indicator features over various time-window lengths (short-term and long-term).
    2. Inclusion of edit distances and seismicity indicators in the same prediction framework
    3. A hierarchical neural network where the first layer has dedicated subnetworks for each time-window length.

Hierarchical Architecture & Features

Seismicity Indicators (per window):
  • T-value
  • Mean magnitude
  • Rate of sqrt-seismic energy
  • Time since first large event
  • Gutenberg-Richter slope & intercept
  • GR fit error
  • Magnitude deficit
  • Coefficient of variation of inter-event times

Preliminary Results

  • Applied to catalogs from New Zealand, Japan, and the Balkans.
  • Average improvement in forecasting accuracy:
    • ~11.8% vs. radial basis functions.
    • ~10.6% vs. NN with only edit distances.
  • The model is in its early stages and shows great promise for future improvement.
Japan Prediction Results

Application to Interregional Seismic Interactions

  • The proposed neural network was then used to analyze interregional seismic interactions.
  • The literature provides strong evidence that distant seismic regions are interconnected through both statistical correlations and physical triggering mechanisms.
  • Large-Scale Stress Transmission & Correlation:
    • Tenenbaum et al.tenenbaum2012earthquake reveals strong correlations between regions in Japan over 1000 km apart.
    • Correlations also exist between seismicity at different depths in distinct areas.takayama1992correlation

Methodology

  • To analyze causal links, Japan and New Zealand were divided into subregions based on centroids over tectonic plate boundaries.
  • Three Experimental Settings:
    1. Cross-region nowcasting: Data from region A to predict current max magnitudes in region B.
    2. Cross-region forecasting: Data from A to predict future max magnitudes in B.
    3. Joint-region forecasting: Data from A + B to predict future max magnitudes in B.

Cross-Region Nowcasting Results

  • We see mixed positive and negative correlations, but the average was statistically positive in both directions (JP → NZ and NZ → JP).
  • For example, predictions for NZ subregion 1 (Lake Pearson) showed a clear positive correlation.

Forecasting & Joint-Region Results

  • Cross-Region Forecasting:
    • Observed a positive average correlation when using JP data to predict for NZ, and vice-versa.
    • Example: NZ subregions 7 (Whanganui) & 8 (Wellington) helped forecast for JP subregion 8 (Nagaoka).
  • Joint-Region Forecasting:
    • On average, adding data from the distant region did not improve local forecasting accuracy.
    • However, some NZ subregions (2, 7, 8) did benefit significantly from JP earthquake data.
    • Thus, the information brought by including data from another region could be in some sense redundant.
  • Interpretation: The connection suggests the earthquake generating processes in both regions might share a common driving force.
  • The lack of improvement could also be due to limitations of the current model, warranting further investigation.

Thesis Contributions: A Multi-faceted Analysis of Edit Distances

  • This thesis analyzes the use of edit distances in earthquake forecasting from several novel perspectives:
  • Efficient Computation:
    • Streamlined edit distance calculation, leveraging parallel computing to make the analysis of large seismic datasets feasible.
  • Theoretical Grounding (during my master's):
    • Formulated the forecasting problem on a solid theoretical basis, using Huke's embedding theorem to justify time-windows and edit distance as a natural metric between them.

Thesis Contributions: A Multi-faceted Analysis of Edit Distances

  • Dynamical Systems Perspective:
    • Conducted a deep analysis of the coupling between earthquakes and other dynamical systems (the Sun and Moon).
    • Results indicate a link between lunar tides to seismicity, with a strong correlation between differential pulls & earthquake magnitudes.
    • Provided strong evidence for a thermal pathway linking solar activity to seismicity, improving forecast accuracy.
  • Bridging Methodologies:
    • Pioneered the integration of our edit distance framework with what is most often used in the literature: statistical indicators.
    • This is done by proposing a novel hierarchical neural network architecture.

Overall Conclusion & Future Outlook

  • This work establishes edit distance as a powerful, theoretically-grounded, and computationally feasible tool for earthquake forecasting.
  • Future Directions (this semester):
    • Development and refinement of the hierarchical neural network to effectively combine diverse precursor data.
      • Inclusion of other statistical indicators.
      • Testing other types of layers (e.g. RNNs, attention layers).
    • Completion of the joint-region analyses, including the Balkan region.

Thank You!