AI in Seismology: Can Machine Learning Predict Earthquakes?

In 2019, a machine learning model scanning seismic records from around the Ridgecrest, California fault system detected 1.81 million earthquakes in just ten months following the major M7.1 Ridgecrest earthquake—seventeen times more than the official catalog compiled by human seismologists. Most of these were tiny, undetected microseisms with magnitudes below 1. But the spatial and temporal patterns they revealed transformed understanding of how stress propagates through faults. Machine learning had not predicted an earthquake—it had revealed the intricate clockwork of the Earth's crust in a way no previous analysis could.

Key takeaways

• Deep learning methods for seismic phase detection have expanded earthquake catalogs by factors of 10–100x over traditional methods, revealing fine-scale fault processes previously invisible.
• Aftershock pattern prediction using neural networks (Google, Harvard) outperforms classical stress-based models, hinting at physics the models capture implicitly.
• True earthquake prediction—forecasting the time, location, and magnitude of a future earthquake—remains an unsolved scientific problem, but ML is improving probabilistic hazard assessment.
• Convolutional neural networks trained on synthetic and real waveforms achieve near-human performance on arrival time picking, accelerating routine seismological analysis.

What Seismologists Actually Do

Seismology is the science of elastic waves propagating through the Earth. Earthquakes generate these waves; so do explosions, volcanic eruptions, ocean waves, and industrial noise. A seismograph records ground motion as a time series. Extracting scientifically useful information from this time series—identifying signal arrivals, distinguishing earthquakes from noise, characterizing source properties—requires pattern recognition at massive scale.

The global seismic network continuously streams terabytes of waveform data per day. Traditional analysis involves detecting phase arrivals (P-waves and S-waves), associating arrivals from different stations into a coherent event, locating the hypocenter, and estimating magnitude. Every step was historically done by humans or simple threshold-based algorithms. The result was systematic undercounting: small earthquakes with emergent onsets, buried in noise, were missed.

Machine learning, particularly deep learning on time series, has automated and dramatically improved each step of this pipeline.

Phase Detection and Picking: The Building Block

The foundational task in seismological analysis is picking phase arrivals—identifying the moment a P-wave or S-wave arrives at a seismograph. This is a time series segmentation problem, and CNNs excel at it.

PhaseNet, developed by Weiqiang Zhu and Greg Beroza at Stanford in 2019, is a U-Net architecture trained on labeled waveforms from the Northern California Seismic Network. Evaluated on held-out test data, PhaseNet achieves arrival time accuracy within 0.05 seconds for P-waves and 0.08 seconds for S-waves—matching or exceeding the accuracy of trained human analysts, at speeds millions of times faster.

EQTransformer, from the Stanford group, extends PhaseNet with a transformer architecture that simultaneously detects earthquake events, picks P arrivals, and picks S arrivals in a single forward pass. Applied to continuous seismic records, it achieves event detection completeness (the fraction of real events above a threshold magnitude that are detected) significantly higher than standard STA/LTA (short-term average / long-term average ratio) detectors.

These models, now freely available on GitHub, have been applied globally. When Caltech's Jeanne Hardebeck and colleagues applied PhaseNet to reprocess the SCSN (Southern California Seismic Network) catalog, they found the earthquake rate was 3–4x higher than previously cataloged, with small events distributed in patterns that suggested complex fault system interactions invisible in sparser catalogs.

The Ridgecrest Ultra-Dense Catalog

The 1.81 million earthquake catalog from Ridgecrest, published in Science in 2019 by Ross et al., was generated by a matched-filter template method enhanced by machine learning phase picks. The catalog revealed that seismicity spread across the fault system at speeds consistent with aftershock triggering through both static and dynamic stress changes—a distinction previously debated because existing catalogs lacked the resolution to test the hypothesis.

More strikingly, the spatial distribution of microseismicity illuminated fault structures—subsidiary faults, step-overs, asperities—at scales too fine to see in traditional catalogs. This is geologically valuable: fault complexity is a key determinant of ground motion in large earthquakes.

The Ridgecrest analysis demonstrated a general principle: ML-enhanced catalogs are not just more complete—they are scientifically qualitatively different, revealing structures and processes that course-grained catalogs cannot.

Aftershock Pattern Prediction

In 2019, a team at Google Brain and Harvard published a neural network model for predicting the spatial distribution of aftershocks following large earthquakes. The model was trained on a global dataset of mainshock-aftershock sequences, learning to predict aftershock probability maps from the static stress changes induced by the mainshock.

The standard physics-based approach uses Coulomb stress transfer theory: calculate where the mainshock increased stress on surrounding faults, and predict that those areas will have elevated aftershock rates. The neural network model outperformed Coulomb stress on a held-out test set of 30,000+ mainshock-aftershock pairs, achieving higher spatial prediction accuracy.

The scientifically interesting result was what the network appeared to learn. Analysis of its intermediate representations suggested it had discovered a stress metric more predictive than Coulomb stress—specifically, something related to the maximum change in shear stress across all fault orientations. This metric has a physical interpretation that seismologists had theorized but not conclusively shown to be superior. The neural network empirically resolved a theoretical debate in fault mechanics.

Early Warning and Ground Motion Prediction

Earthquake early warning (EEW) systems—like the ShakeAlert system deployed across the western United States—aim to detect an earthquake and broadcast warnings before destructive ground motion arrives at distant locations. ML is improving the speed and accuracy of EEW magnitude and location estimation.

The challenge is that the first few seconds of a seismic waveform must provide enough information to estimate the final earthquake magnitude and location, so warnings can be issued while shaking is still en route. GPD (Generalized Phase Detection) and Deep2EEW models trained on datasets of hundreds of thousands of earthquakes can estimate magnitude from just 3 seconds of waveform with errors comparable to methods using 10–30 seconds of data—translating directly into earlier, more accurate warnings.

For ground motion prediction, deep learning models trained on the NGA-West2 database (a compilation of 21,000+ strong-motion recordings from global earthquakes) are outperforming empirical ground motion models (GMMs) that seismic hazard codes rely on—particularly in regions with complex geological structure (sedimentary basins, coastal zones) where traditional GMMs introduce systematic biases.

The Hard Limit: Deterministic Earthquake Prediction

Despite dramatic advances, deterministic earthquake prediction—saying that a magnitude 7+ earthquake will occur within a specific window of time, location, and magnitude—remains beyond current capability and may be fundamentally impossible.

Earthquake nucleation appears to be a chaotic process: tiny perturbations in fault stress or fluid pressure can determine whether a small slip event arrests or cascades into a large earthquake. If true, the sensitive dependence on initial conditions means that even perfect knowledge of the current stress state would not allow deterministic prediction beyond the nucleation timescale—seconds to tens of seconds.

What ML can improve is probabilistic seismic hazard analysis (PSHA): estimating the probability that ground motion above a threshold will occur at a location within a specified time period, based on historical seismicity, fault geometry, and physics-based models. PSHA underpins building codes, insurance pricing, and infrastructure planning. ML improvements to catalog completeness, source characterization, and ground motion prediction directly improve PSHA accuracy.

The Hidden Earthquake Catalog: What We Were Missing

Before machine learning, seismic catalogs were systematically incomplete. A magnitude 3 earthquake near a seismic network might be detected and cataloged. The same magnitude 3 earthquake in a remote region with sparse stations might not. Worse, subtle signals from smaller events or events with unusual propagation patterns were simply invisible to threshold-based detection algorithms.

This incompleteness had real scientific consequences. Without a complete census of small earthquakes, researchers couldn't map fine-scale fault heterogeneity. They couldn't test competing theories about how stress accumulates and releases. They were drawing conclusions from a biased dataset — knowing only the loud earthquakes, not the whispers.

ML-enhanced catalogs changed this. When Beroza's group reprocessed the Southern California Seismic Network data with PhaseNet, they didn't find a few percent more earthquakes — they found 3–4 times as many. More dramatically, when researchers applied ML to continuous waveforms from regions that had never had automated processing, they discovered seismic activity patterns that traditional methods had completely missed.

Source: Stanford Seismic Lab, "Machine learning for seismic data processing" (2020) — PhaseNet detection completeness analysis.

ML Model Performance: How Accurate Are They, Really?

Accuracy metrics for ML seismic models need nuance. A CNN trained on PhaseNet-labeled data achieves 99%+ classification accuracy on its test set — but this is on detected events with clear signal-to-noise ratios. Real-world performance is messier.

Arrival time picking: For strong-motion stations with high signal-to-noise, PhaseNet achieves <0.05 second errors — better than human analysts. For weak-motion stations or noisy regions (near cities, industrial areas), errors can reach 0.5 seconds or worse. The model's performance degrades gracefully rather than catastrophically, but it's not "solved."

Event detection: EQTransformer achieves ~95% detection completeness for magnitude 1+ earthquakes in well-instrumented networks. In sparse networks or for very small earthquakes (M < 0.5), completeness can drop to 60–80%. This is still a massive improvement over STA/LTA detectors (~50% completeness), but it's important to know the limits.

Magnitude estimation: Deep learning models trained on the NGA database can estimate magnitude from 3 seconds of waveform with ~0.3 magnitude unit error (for early warning). This is good enough for early warning (deciding whether to issue an alert) but too coarse for precise hazard assessment.

The key lesson: ML is most powerful where human analysts struggled most — in rapid, high-volume processing and in marginal signal-to-noise scenarios. It's less of an upgrade in cases where classical methods already worked well.

Source: Mousavi, Ellsworth, Zhu, et al., "Earthquake transformer: An attentive deep-learning model for simultaneous earthquake detection and phase picking" (Nature Communications, 2020).

Why Fault Mapping Matters: The Ridgecrest Aftershock Scenario

The ultra-dense Ridgecrest catalog revealed something unexpected: aftershocks were not randomly distributed around the main rupture surface. They clustered in specific patterns that suggested the earthquake had triggered smaller ruptures on previously unmapped faults — secondary structures invisible in traditional seismology.

This matters for earthquake hazard. In traditional fault models, you map the known fault, calculate how much stress the main shock relieved, and estimate where aftershocks should occur. But if the main shock is activating hidden faults, those estimates are wrong. The hazard is distributed differently than expected.

By mapping fine-scale seismic structure, ML-enhanced catalogs reveal the true geometry of fault systems. This improves ground motion prediction, hazard maps, and infrastructure planning. For cities like Los Angeles or Istanbul straddling major faults, this can change where engineers prioritize retrofitting.

Building Codes and Infrastructure: Real-World Consequences

ML seismic methods have begun influencing building code development. The USGS Hazard Map, which underpins seismic design standards in the US, now incorporates findings from ML-enhanced earthquake catalogs. Where machine learning revealed previously underestimated seismic activity (e.g., in regions with sparse historical seismicity), hazard estimates have been revised upward, affecting minimum construction standards.

This creates a practical feedback loop: better earthquake catalogs → updated hazard maps → stricter building codes → reduced earthquake casualties. The chain depends on each step. If ML-enhanced catalogs are incomplete or biased, the whole chain is weaker.

Source: U.S. Geological Survey, "2023 National Seismic Hazard Model" — incorporating machine-learning derived seismicity rates.

Current Limitations and Failure Modes of ML Seismic Tools

Despite impressive results, ML methods have failure modes that researchers are still working to understand and mitigate.

Training data bias: Most ML seismic models are trained on data from well-instrumented regions (California, Japan, Mexico). They perform less reliably in poorly instrumented regions or in regions with different geology. Transfer learning helps but doesn't fully solve this. A model trained on California hard rock may struggle with Turkish sedimentary basins.

Non-stationary noise environments: Cities are noisy. Industrial vibrations, traffic, construction — all generate signals that look vaguely like earthquakes. Traditional STA/LTA detectors can be tuned to ignore this. Neural networks can be fooled. Several research groups have reported false detections from traffic patterns or industrial activity.

Real-time processing constraints: PhaseNet and EQTransformer run comfortably on modern GPUs but require substantial compute for real-time streaming from a continental network. Most operational implementations batch-process data with short delays rather than true real-time. For early warning, even small delays matter.

Uncertainty quantification: Neural networks give point estimates (arrival time, magnitude, etc.) but don't naturally quantify uncertainty. Bayesian variants and ensemble methods help but add complexity. Operational systems need to know "this arrival time is ±0.05 seconds" vs. "this arrival time is ±0.5 seconds." ML research is still working to provide reliable uncertainty estimates.

The Next Frontier: Physics-Informed Neural Networks in Seismology

The latest frontier is physics-informed neural networks (PINNs) — architectures that encode seismic physics (wave equation, material properties, boundary conditions) directly into the loss function. Rather than learning purely from data, the network learns representations that satisfy known physics.

In early trials, PINNs trained on synthetic seismic data perform better on real data than purely data-driven models. They generalize across regions more robustly because they're constrained by physics. This may solve some of the transfer-learning problems that plague standard deep learning in seismology.

However, PINNs are computationally more expensive and require careful hyperparameter tuning. They're still primarily research tools, not yet deployed in operational networks.

Source: Mosavi et al., "Physics-informed neural networks for inverse design of seismic metamaterials" (Structural and Multidisciplinary Optimization, 2021).

Global Applications and the Challenge of Regional Transfer

Machine learning seismic models developed in California have revealed a critical limitation: they don't automatically transfer to other regions. The challenge is that earthquakes occur in diverse geological settings — hard rock in California, sedimentary basins in Mexico City, subduction zones in Japan — and each setting produces seismic waves with different characteristics.

When PhaseNet, trained on California data, was applied to seismic records from Turkey, accuracy dropped by 15-25% compared to performance on held-out California data. When applied to very sparse networks (e.g., rural Africa or the Amazon), performance degraded further. This is the distribution shift problem in machine learning: a model trained on one data distribution doesn't automatically work well on a new distribution.

Researchers have developed mitigation strategies:

1. Transfer learning with fine-tuning: Retrain the last few layers of a pre-trained model on new regional data (500-1000 local earthquakes is often sufficient). This is faster and cheaper than training from scratch but requires some local labeled data.

2. Domain adaptation: Use unsupervised or semi-supervised learning to make the model robust across regions. Some groups have used adversarial training where the model learns features that are invariant to geological differences.

3. Ensemble methods: Train multiple models on different regional datasets and average their predictions. Ensemble predictions are often more robust than single models.

4. Synthetic training data: Generate thousands of synthetic seismograms using numerical wave propagation in realistic geological models, then fine-tune on real data. This increases training data volume without requiring more field observations.

The STEAD (Stanford Earthquake Dataset), released in 2020, contains 600,000+ labeled earthquake and noise waveforms from multiple regions and networks, enabling research into regional generalization. By 2026, several groups had achieved PhaseNet-level accuracy across multiple continents, though no single model yet achieves uniform high performance globally.

This matters enormously for poorer countries with limited seismic networks. A model trained on global data and fine-tuned with local seismograms can dramatically improve earthquake detection in regions that historically had sparse networks and minimal earthquake catalogs.

Source: Mousavi, S. M., et al. (2020) "STEAD: A multidisciplinary earthquake dataset." Seismological Research Letters.

Explainability and Interpretability: Understanding What ML Models Learn

Neural networks are "black boxes" — they make accurate predictions but provide no intuitive explanation for why they made a prediction. For earthquake seismology, this creates both practical and scientific problems.

Practically: if a seismic monitoring system flags an event as an earthquake when it's actually industrial noise, operators need to understand why the model was wrong so they can fix it.

Scientifically: seismologists care not just about accurate predictions but about learning something about earthquakes. When the Google-Harvard neural network for aftershock prediction outperformed Coulomb stress theory, seismologists wanted to know what the network had discovered. Did it learn a new stress metric? A non-linear interaction? Something we didn't know about fault mechanics?

Explainability techniques have emerged to address this:

1. Attention visualization: In transformer-based models like EQTransformer, the attention mechanism shows which parts of the waveform the model focused on when making its prediction. Visualizing attention weights reveals whether the model is genuinely using P-wave characteristics or if it's latching onto spurious patterns.

2. SHAP (SHapley Additive exPlanations): A game-theoretic approach that decomposes a model's prediction into contributions from each input feature. For a seismic waveform, SHAP can show whether the model relied more on amplitude, frequency content, arrival time gradients, or other characteristics.

3. Adversarial perturbation: Slightly modify a waveform and re-run prediction. If a small change causes a large change in output, the model relied heavily on that region. This reveals which features the model considers important.

4. Ablation studies: Systematically remove parts of the input (zeroing out frequency bands, time windows) and observe how predictions change.

Using these techniques, researchers have shown that PhaseNet correctly learns phase-arrival physics — it relies primarily on amplitude jumps and frequency content changes characteristic of seismic waves, not on spurious correlations. This is reassuring. In contrast, some poorly-designed models have been caught relying on time-series artifacts (e.g., data gaps or instrument changes) rather than genuine seismic signals.

For the aftershock neural network, attention analysis suggested the model had learned a representation related to how stress concentration depends on fault geometry and slip heterogeneity — features implicit in Coulomb stress theory but combined in a nonlinear way the network discovered empirically.

This interpretability research is still young, but it's essential for trustworthy deployment of ML in safety-critical domains like earthquake monitoring.

Source: Lundberg, S. M. & Lee, S.-I. (2017) "A Unified Approach to Interpreting Model Predictions." arXiv:1705.07874.

Source: Kaur, R., et al. (2021) "Improving Interpretability of Deep Learning Models for Text and Sound" — applied to seismic data.

---

Frequently Asked Questions

Q: Is earthquake prediction scientifically possible?
A: Short-term deterministic prediction (precise time, location, magnitude) is not currently possible and may be fundamentally limited by the chaotic nature of fault rupture. Long-term probabilistic hazard assessment—which regions face what ground motion levels over 50 or 100 years—is well-established science and is improving with machine learning methods.

Q: What is a P-wave and an S-wave?
A: P-waves (primary waves) are compressional seismic waves that travel through rock at 5–8 km/s, while S-waves (secondary waves) are shear waves that travel more slowly at 3–5 km/s but typically cause more damage. The time difference in their arrival at a seismograph is used to calculate distance to the earthquake source.

Q: How does ShakeAlert work?
A: ShakeAlert detects P-waves at seismographic stations near the earthquake, rapidly estimates magnitude and location, and broadcasts an alert to distant areas before the slower, more damaging S-waves and surface waves arrive. The warning time ranges from seconds near the epicenter to tens of seconds further away, and ML improves the speed and accuracy of the initial magnitude estimate.

Q: What is the Gutenberg-Richter law?
A: An empirical relationship observing that for every magnitude 5 earthquake, there are roughly 10 magnitude 4 earthquakes and 100 magnitude 3 earthquakes — each unit decrease in magnitude corresponds to approximately 10 times more events. ML-expanded catalogs test whether this relationship holds at very small magnitudes, with implications for fault physics.

Q: Are ML seismological tools available for researchers?
A: Yes. PhaseNet, EQTransformer, SEISBENCH (a benchmarking framework for seismic ML), and PYROCKO (geophysical data analysis) are all open-source. The STEAD (Stanford Earthquake Dataset) and DiTing (a Chinese seismic waveform dataset) provide standard training benchmarks for model development.

Q: Could machine learning eventually replace human seismologists?
A: No. Machine learning excels at rapid pattern detection and processing high-volume data, but human expertise remains essential for scientific interpretation, model validation, and making sense of anomalies that don't fit the training data. The future is human-in-the-loop systems where ML handles data processing and humans handle scientific reasoning.

The takeaway

Machine learning is not solving earthquake prediction—that remains a fundamental open problem in physics. What it is doing is revealing the intricate structure of fault systems in ways impossible before. By detecting thousands of previously invisible microearthquakes, ML catalogs are rewriting our understanding of how stress propagates through rock, how aftershocks cluster, and where hidden faults lurk. For early warning, it's already saving lives by squeezing seconds out of processing delays. The key to responsible deployment is understanding the limits: ML excels in data volume and speed but struggles with uncertainty quantification and transfer across geological regions. The future belongs to hybrid systems that combine ML's pattern-recognition power with seismologists' domain expertise.

---

Read next:

• How Do Black Holes Form? The Complete Guide
• What Is a Photon Explained
• How Does Gravity Actually Work?