Abnormal low-magnitude seismicity preceding large-magnitude earthquakes

The study addresses whether low-magnitude seismicity contains reliable, short-term precursory signals of impending large earthquakes and how these can be detected and interpreted. Forecasting the timing, size, and location of major earthquakes remains a formidable challenge, as many proposed geophysical, geochemical, and biological precursors are debated or unreliable. With the advent of high-quality earthquake catalogs and advanced machine-learning techniques, the authors hypothesize that nonlinear, multivariate patterns in low-magnitude seismicity may reveal regional tectonic unrest weeks to months before large events. They apply and test this idea on two major cases: the 2019 Ridgecrest sequence (California) and the 2018 Anchorage earthquake (Alaska), exploring whether anomalous low-magnitude seismicity (M between 1 and 6) precedes these events similarly to prior large earthquakes in the same regions.

Prior work has investigated numerous earthquake precursors, including thermal infrared anomalies, acoustic and acoustic-gravity waves, groundwater level and composition changes, gas emissions (e.g., radon), electromagnetic and ionospheric perturbations, crustal deformation, slow slip events, and even animal behavior. However, many such signals remain contentious. Statistical indicators like changes in the Gutenberg–Richter b-value and other anomalies in low-magnitude seismicity have been explored, but clear, generalizable patterns are elusive due to the complexity of seismic spatiotemporal distributions. Recent advances in machine learning, both supervised and unsupervised, have shown promise in identifying hidden, nonlinear patterns in laboratory and analog fault systems and in aftershock forecasting, suggesting potential applicability to real earthquake precursors in catalog data. The study builds upon this body of work by applying a multivariate, supervised approach to detect robust precursory patterns in regional low-magnitude seismicity.

Data: The authors retrieved USGS earthquake catalog data for M ≥ 1.0 events from 1989-01-01 to 2020-12-31 for Southern California (lat 32.5°–38°, lon −124° to −112°) and Southcentral Alaska (lat 56°–65°, lon −159° to −143°). Data from 1989–2012 were used for training; 2013 onward for testing. Extracted fields include occurrence time, depth, magnitude, and epicentral coordinates. Machine learning model: The approach targets detection of anomalous low-magnitude seismicity (MMin=1, MMax=6) preceding large-magnitude events (MLME ≥ 6.4). Spatiotemporal nodes are defined with circular regions of radius R=120 km and 1-year backward sliding windows stepping daily over a T=2-year span. For training, nodes include epicenters and times of historical large events in the two regions (Southern California: 1992 M7.3 Landers, 1994 M6.7 Northridge, 1999 M7.1 Hector Mine, 2003 M6.6 San Simeon; Southcentral Alaska: 1999 M7.0, 1999 M6.4, 2000 M6.5, 2001 M6.9 (Kodiak), 2001 M6.7 (Southern Alaska), 2002 M6.6 and 2002 M7.9 Central Alaska) plus up to N=50 random nodes distributed regionally. This selection is repeated 200 times to form ensembles. Features: For each node’s windowed catalog, five standardized time series are computed: standard deviations of inter-event time (σIET), depth (σD), latitude (σLAT), longitude (σLON), and magnitude (σMAG). Labeling: Time steps within Tunrest days (nominally Tunrest=30 days in main tests; also 10–100 days in sensitivity tests) before large events at their epicenters are labeled as unrest (class 1); all other steps at other nodes and times are quiescence (class 0). Model training and inference: For each ensemble, a random forest (RF) with 500 trees is trained, drawing two features per iteration, using 10-fold cross-validation (3 repeats). The degree/probability of unrest Pun is defined as the fraction of trees classifying a time step as unrest; it can be interpreted as the probability of MLME ≥ 6.4 within the next Tunrest days (30 days in main tests). For testing (2013+), the RF ensemble is run at specified target locations and on regional grids (0.1°×0.1°) to generate daily Pun at epicenters and spatial probability maps. Parametric analysis: Sensitivity tests varied MMin (≥1 to ≥1.5), MMax (≤6), Tunrest (10–100 days), and R (20–150 km), and trained with individual features to assess importance and robustness. Finite element models: To interpret mechanisms, 2D (and additional 3D) finite element solid mechanics models (COMSOL v6.2) simulate a 2000×2000 km elastic domain hosting one large fault (LF, length 850 km) and a network of 15 small faults (40–97 km). Material properties: host rock Young’s modulus 30 GPa, ν=0.25, density 2300 kg/m³; small faults with stiffness 0.03–0.92 GPa (constant); LF stiffness varied between 10 GPa and 0.01 GPa. Two scenarios: (1) Progressive horizontal compression from 10 to 20 MPa with constant LF stiffness; (2) Constant compression at 20 MPa with progressive decrease of LF stiffness. Von Mises stress accumulation on small faults is analyzed to assess regional stress redistribution as failure approaches.

Epicentral time series: At Ridgecrest, Pun remained low with minor spikes (<~5%) until ~40 days before the sequence, then mean Pun rose to ~20–25% and maximum to ~75%, remaining elevated to the first M6.4 event; after the M6.4 event, maximum Pun increased to 90% before decaying to background (≤5%) roughly three months after. At Anchorage, maximum Pun increased abruptly to ~80% around three months before the M7.1 event; mean and maximum Pun reached ~35% and ~85% in the days just before. A lower-level anomalous phase ~700 to ~300 days prior was identified as background spatial noise upon area analysis. Following the Anchorage mainshock, Pun rose up to ~99% over three months, then declined below ~20% by ~170 days and to ~5% by ~250 days. Regional maps and area metrics: Precursory anomalies were spatially widespread, emerging first in select zones (e.g., southern San Joaquin Valley, California–Nevada–Arizona border) ~90–60 days before Ridgecrest, then spreading to multiple fault systems including the eventual epicentral area by ~30 days prior. In Alaska, unrest emerged center-northeast ~80 days and south-southeast ~10 days before Anchorage. The area with Pun ≥ 50% increased by up to an order of magnitude during the three months preceding the main events: in Southern California from ~2% to ~17%, and in Southcentral Alaska from ~2% to ~22%. Catalog quality tests: Analyses of event magnitude versus order and moving-window Gutenberg–Richter fits (mGR for M=1–3) show no precursory decrease in mGR and no pre-event degradation in catalog completeness; incompleteness is only evident post-mainshock, indicating the precursory anomalies are not artifacts of changing detection thresholds. Parametric results: Findings are robust across hyperparameters. Precursory patterns are most pronounced with broad low-magnitude definitions (MMin=1, MMax=6), short Tunrest (≤30 days), and larger regions of interest (R ≥ 120 km). Patterns largely vanish for MMin ≥ 1.5, implicating very low-magnitude events in the signal. Among features, depth variability contributes most; magnitude variability is least informative. Mechanistic modeling: FEM results show that simple regional loading (increasing compression) raises stress across all small faults uniformly, which does not explain nonlinear anomalies. In contrast, decreasing stiffness of a nearby large fault under constant loading produces dissimilar, nonlinear stress changes (increases at some small faults and decreases at others), consistent with the observed regional, uneven precursory seismicity. This supports a mechanism of stiffness reduction due to elevated pore fluid pressure and damage in large fault segments approaching failure.

The results demonstrate that low-magnitude seismicity exhibits multivariate, nonlinear anomalies weeks to months prior to large earthquakes and that these anomalies are regionally distributed rather than confined strictly to the epicenter. By validating catalog completeness and GR behavior, the study attributes anomalies to genuine tectonic processes. FEM modeling links the observed patterns to a reduction in stiffness of major fault segments—likely due to increased pore fluid pressure and damage—causing uneven regional stress redistribution that differentially pushes some small faults toward failure while relieving others. These findings address the research question by showing that machine learning can extract robust precursory signals from routine catalog data, offering probabilistic indications of heightened regional unrest on actionable timescales (days to months). The spatial probability maps, while broad (tens to hundreds of kilometers), constrain potential locations and complement timing and magnitude-threshold information. Compared with prior studies reporting both shorter (minutes) and longer (years) lead times, the identified weeks-to-months window adds a consistent intermediate-term layer of forecasting potential and suggests a multifaceted preparatory process operating across scales. The approach could enhance operational monitoring by flagging elevated unrest levels to prompt targeted deployments and preparedness measures.

The study introduces a multivariate, random forest-based method that detects abnormal low-magnitude seismicity as a precursor to major earthquakes, validated on the 2019 Ridgecrest sequence and the 2018 Anchorage event. It shows that regional tectonic unrest emerges weeks to months beforehand, covering increasing fractions of the affected regions, and that mechanistic FEM modeling can explain these observations through stiffness reductions in major faults driven by fluid pressure and damage. The approach relies solely on routinely archived catalog parameters and can be integrated into near-real-time surveillance to inform alert-level strategies. Future work should extend the framework to additional regions with complete catalogs, incorporate geodetic observations to refine mechanistic understanding and spatial constraints, and retrain models as new large events occur to maintain robustness and adaptability.

The algorithm requires sufficient low-magnitude seismicity; regions with sparse activity cannot be reliably assessed. Training depends on an adequate number of past large events (M ≥ 6.4) in the target region to capture representative precursory patterns. The method assumes future anomalies will resemble those seen historically, necessitating periodic retraining after major events. Spatial forecasts delineate broad unrest areas rather than pinpoint epicenters. Catalog completeness variations immediately after large events can affect post-event signals. Application to new regions requires region-specific training with local historical seismicity.