Neural Network Learns to Predict Earthquake Shaking Across the Frequency Spectrum

Earthquake engineers have long relied on equations carved out by hand to predict how hard the ground will shake during an earthquake. Now, a team of researchers has handed that job to an artificial neural network, training it on one of the largest collections of earthquake recordings ever assembled. The result, published in the Bulletin of Earthquake Engineering, is a data-driven ground-motion model that predicts the full frequency content of shaking from shallow crustal earthquakes across the western United States, and it does so without being told in advance what mathematical shape the answer should take.

The study, led by Farhad Sedaghati of Aon Impact Forecasting together with Shahram Pezeshk and Mehran Davatgari-Tafreshi of the University of Memphis, focuses on a quantity called the Effective Amplitude Spectrum, or EAS. Unlike the response spectra most commonly used in building design, which describe how an idealized oscillator reacts to shaking, the EAS is derived directly from the Fourier amplitude spectrum of the recorded motion. That gives it a direct physical interpretation: it reflects how energy is distributed across frequencies by the earthquake source, modified by the path the waves travel and by the soil conditions beneath the recording station. Engineers can also feed EAS predictions into random vibration theory to estimate response spectra quickly, without running time-consuming simulations.

To train their model, the researchers drew on the NGA-West3 database, a large and uniformly processed collection of three-component accelerograms from shallow crustal earthquakes in active tectonic regions. The raw compilation contained 175,284 records, but the team applied a stringent sequence of quality-control filters before a single neuron was trained. Records were excluded if the earthquake was smaller than magnitude 3.0, if the hypocenter was deeper than 30 kilometers, if the rupture distance exceeded 300 kilometers, or if metadata such as fault dip or shear-wave velocity were missing or implausible. Sensors buried more than two meters below the ground surface were rejected, as were stations classified as non-free-field. After all screening steps, 50,639 high-quality records remained, roughly 90 percent of them from California, with the rest contributed by Alaska, Idaho, Nevada, Utah, Washington, and Wyoming.

The neural network itself is deliberately compact. It takes six explanatory variables as input: moment magnitude, rupture distance, fault dip, depth to the top of the rupture, the time-averaged shear-wave velocity in the upper 30 meters of the site, and the depth to a stiff 2.5 kilometers-per-second shear-wave velocity horizon that characterizes sedimentary basins. From these six numbers, the network predicts the natural logarithm of EAS at 22 discrete frequencies spanning 0.05 to about 33 hertz, all at once. This multi-output design proved crucial. When the researchers initially tried predicting each frequency with a separate single-output model, the resulting spectra were jagged and unstable. Predicting all frequencies simultaneously through shared hidden layers allowed the network to learn cross-frequency patterns directly from the data, producing smooth, physically plausible spectra.

One of the thorniest practical problems was missing data. A recording may contain reliable spectral information at some frequencies but not others, because signal-to-noise ratios degrade at both the lowest and highest frequencies. Rather than discarding every record with a single gap, the team used a masked loss function during training. Missing values, represented as not-a-number entries, simply do not contribute to the error that the network tries to minimize, so partially observed spectra still teach the model whatever they can. This masked approach substantially increased the effective amount of data available, particularly at the frequency extremes where coverage is thinnest.

Avoiding inflated estimates of accuracy required equally careful design of the evaluation strategy. Recordings from the same earthquake share source characteristics, so splitting the data randomly would let the network effectively memorize events it was supposed to predict. The researchers therefore partitioned the dataset at the event level, keeping all recordings of a given earthquake together in the same fold. Hyperparameters were selected through ten-fold event-grouped cross-validation, and roughly five percent of events were held out entirely as an independent evaluation set, untouched during architecture selection, preprocessing, or training decisions. After that out-of-event assessment was complete, the final network was retrained on the full retained dataset so that every available earthquake contributed to the distributed model. The winning architecture turned out to be modest: two hidden layers with 32 and 16 neurons, a batch size of 64, and a learning rate of ten to the minus two.

The predictions that emerged show textbook seismological behavior, despite the fact that no functional form was ever imposed. Spectral amplitudes rise monotonically with magnitude and fall with distance, with high frequencies decaying faster than low ones as waves scatter and lose energy in the crust. At low frequencies the spectra steepen toward a corner frequency in a manner consistent with the classical omega-square source model, and at high frequencies they roll off in the way associated with near-surface damping. The team is careful to note that these are qualitative consistencies, not quantitative estimates of source parameters such as corner frequency or kappa, and that predictions above magnitude 7.36, the largest event in the training data, are extrapolations presented for illustration rather than validated forecasts.

Beyond median predictions, the researchers quantified aleatory variability, the irreducible randomness in ground motion that remains after source, path, and site effects are accounted for. Using a sequential mixed-effects decomposition, they separated total residuals into between-event, site-to-site, and event-site-corrected components at each of the 22 frequencies. The standard deviations show clear frequency dependence, increasing toward higher frequencies where ground motion is most sensitive to record-specific details, and they fall within the range reported by existing empirical EAS models. This decomposition matters for practice: in conventional probabilistic seismic hazard analysis the total variability is used directly, but for site-specific studies the explicit separation allows engineers to condition on measured site behavior and reduce uncertainty.

How does the machine stack up against the human-crafted equations it joins? The team compared their model against three established empirical EAS ground-motion models, Bora and colleagues from 2019, Bayless and Abrahamson from 2019, and Campbell and Bozorgnia from 2025, all built on the earlier NGA-West2 database. Across representative scenarios, the neural network’s median predictions track the empirical models closely in both spectral shape and scaling trends, and its variability estimates sit within the published range. The authors are candid that these comparisons demonstrate physical plausibility rather than superiority, since a rigorous head-to-head test would require fitting a parametric model to identical data with identical partitions, something they flag as important future work. Interestingly, the network also reproduced the observed cross-frequency correlation structure of the data with a Pearson correlation of 0.964 between observed and predicted correlation coefficients.

The researchers have released the trained model, its preprocessing scaler, and example code on GitHub, making the tool immediately usable by hazard analysts. Because the network produces broadband Fourier spectra in a fraction of a second, it is well suited to applications that demand thousands or millions of repeated ground-motion evaluations, such as probabilistic seismic hazard analysis and simulation-based risk studies. The model is strictly valid within the range of its training data, magnitudes from about 3.0 to 7.36, distances up to 300 kilometers, and sites from soft soil to hard rock, and it is calibrated specifically for western United States tectonics. Future directions include explainable artificial intelligence techniques to reveal which predictors matter most at each frequency, and physics-informed training that would embed source models such as Brune’s spectrum directly into the learning objective, potentially improving extrapolation to the largest, rarest earthquakes that engineers care about most.

Subject of Research: A data-driven neural network ground-motion model for Fourier amplitude spectra of shallow crustal earthquakes using the NGA-West3 database

Article Title: A data-driven ground-motion model for Fourier amplitude spectra of shallow crustal earthquakes using the NGA-West3 database

Article References: Sedaghati, F., Pezeshk, S., & Davatgari-Tafreshi, M. (2026). A data-driven ground-motion model for Fourier amplitude spectra of shallow crustal earthquakes using the NGA-West3 database. Bulletin of Earthquake Engineering. https://doi.org/10.1007/s10518-026-02708-1

Image Credits: AI Generated

DOI: 10.1007/s10518-026-02708-1

Keywords: ground-motion model, effective amplitude spectrum, NGA-West3, artificial neural network, machine learning, Fourier amplitude spectrum, seismic hazard analysis, shallow crustal earthquakes, aleatory variability, random vibration theory, western United States, earthquake engineering