When linear inversion fails:Neural-network optimization for sparse-ray travel-time tomography of a volcanic edifice

In this study,we present an artificial neural network(ANN)-based approach for travel-time tomography of a volcanic edifice under sparse-ray coverage.We employ ray tracing to simulate the propagation of seismic waves through the heterogeneous medium of a volcanic edifice,and an inverse modeling algorithm that uses an ANN to estimate the velocity structure from the“observed”travel-time data.The performance of the approach is evaluated through a 2-dimensional numerical study that simulates i)an active source seismic experiment with a few(explosive)sources placed on one side of the edifice and a dense line of receivers placed on the other side,and ii)earthquakes located inside the edifice with receivers placed on both sides of the edifice.The results are compared with those obtained from conventional damped linear inversion.The average Root Mean Square Error(RMSE)between the input and output models is approximately 0.03 km/s for the ANN inversions,whereas it is about 0.4 km/s for the linear inversions,demonstrating that the ANN-based approach outperforms the classical approach,particularly in situations with sparse ray coverage.Our study emphasizes the advantages of employing a relatively simple ANN architecture in conjunction with second-order optimizers to minimize the loss function.Compared to using first-order optimizers,our ANN architecture shows a~25%reduction in RMSE.The ANN-based approach is computationally efficient.We observed that even though the ANN is trained based on completely random velocity models,it is still capable of resolving previously unseen anomalous structures within the edifice with about 5%anomalous discrepancies,making it a potentially valuable tool for the detection of low velocity anomalies related to magmatic intrusions or mush.

A study on small magnitude seismic phase identification using 1D deep residual neural network

Reliable seismic phase identification is often challenging especially in the circumstances of low-magnitude events or poor signal-to-noise ratio.With improved seismometers and better global coverage,a sharp increase in the volume of recorded seismic data has been achieved.This makes handling seismic data rather daunting by using traditional approaches and therefore fuels the need for more robust and reliable methods.In this study,we develop 1D deep Residual Neural Network(ResNet),for tackling the problem of seismic signal detection and phase identification.This method is trained and tested on the dataset recorded by the Southern California Seismic Network.Results demonstrate that the proposed method can achieve robust performance for the detection of seismic signals and the identification of seismic phases.Compared to previously proposed deep learning methods,the introduced framework achieves around 4%improvement in earthquake detection and a slightly better performance in seismic phase identification on the dataset recorded by Southern California Earthquake Data Center.The model generalizability is also tested further on the STanford EArthquake Dataset.In addition,the experimental result on the same subset of the STanford EArthquake Dataset,when masked by different noise levels,demonstrates the model’s robustness in identifying the seismic phases of small magnitude.

PolarCAP – A deep learning approach for first motion polarity classification of earthquake waveforms

The polarity of first P-wave arrivals plays a significant role in the effective determination of focal mechanisms specially for smaller earthquakes.Manual estimation of polarities is not only time-consuming but also prone to human errors.This warrants a need for an automated algorithm for first motion polarity determination.We present a deep learning model-PolarCAP that uses an autoencoder architecture to identify first-motion polarities of earth-quake waveforms.PolarCAP is trained in a supervised fashion using more than 130,000 labelled traces from the Italian seismic dataset(INSTANCE)and is cross-validated on 22,000 traces to choose the most optimal set of hyperparameters.We obtain an accuracy of 0.98 on a completely unseen test dataset of almost 33,000 traces.Furthermore,we check the model generalizability by testing it on the datasets provided by previous works and show that our model achieves a higher recall on both positive and negative polarities.