Pith. sign in

REVIEW 4 major objections 7 minor 42 references

Generalised Time-Series Analysis of Fault Mechanics Using Explainable AI

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A time-delay neural network trained on acoustic emissions from granite predicts when and how strongly the rock will fail, while its shifting input weights trace the three stages of fault growth.

desk verdict A useful feature-engineering demonstration that oversells itself as a predictive tool; the target-dependent normalization and one-sample-per-pressure validation undermine the generalization claim. read the letter →

arxiv 2505.21312 v1 pith:UUS67J3F submitted 2025-05-27 physics.geo-ph

Thomas King , Sergio Vinciguerra This is my paper
classification physics.geo-ph
keywords acousticemissionstimedelayneuralnetworkfaultnucleationandgrowthtriaxialrockdeformationexplainablemachinelearningpeakGutenberg-Richterb-valuefractaldimension
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a neural network can serve as a mechanical-state monitor for deforming rock: given a stream of acoustic-emission measurements from a granite sample being crushed under triaxial pressure, it can reproduce the stress-strain history and flag the approach of failure. The network is a time-delay neural network whose architecture is tuned by a genetic algorithm, and its inputs mix waveform-derived features (peak delay, mean free path, coda Q) with event-statistics features (event rate, b-value, fractal dimension). The paper's specific claim is that this model predicts stress-drop timing and magnitude at confining pressures from 5 to 40 MPa, covering failure modes from axial splitting to shear localisation. The explanatory twist is that the network's input weights are time-dependent: it leans on waveform features early, switches to event-statistics features as damage localises, and returns to waveform features near coalescence, which the authors read as the physical fingerprint of fault development. If correct, this gives a data-driven way to recognise which stage of fault growth a rock mass is in, with potential application to earthquake hazard and structural monitoring.

What carries the argument

The engine is a Time Delay Neural Network (TDNN), a neural network whose convolution filters slide along the time axis, letting it classify shift-invariant temporal patterns in sequential data. A genetic algorithm searches the network's architecture (number of hidden layers, neurons, and time delays), and a moving-window sensitivity analysis tracks the weights assigned to each acoustic-emission input as the experiment progresses, separating waveform-derived from occurrence-derived contributions. This weight trajectory is the mechanism that carries the explanatory claim: it is what the authors use to identify the three fault-growth phases and to argue the model is physically interpretable rather than a black box.

What would settle it

Retrain the same network on features computed in fixed absolute time windows, with the time axis expressed in elapsed time rather than strain at failure, and hold out one confining pressure as before; if the held-out prediction error does not worsen materially, the current accuracy is not an artifact of target leakage, whereas a large degradation would show that the normalisation, not the acoustic emissions, carried the forecast.

Watch

Extended reading notes

Core claim

The central claim is that a Time Delay Neural Network optimised by a genetic algorithm can learn the evolving mechanical state of Alzo granite from multi-parameter acoustic-emission data and predict the timing and magnitude of stress drops at failure across confining pressures of 5, 10, 20, and 40 MPa. The network combines waveform-derived attributes (relative peak delay, mean free path, and total coda Q, estimated through a diffusion model of wave propagation) with occurrence-derived metrics (AE event rate, b-value, and spatial fractal dimension) and is validated by training on three pressure conditions and testing on the held-out fourth. A second claim is that the network is explainable through a moving-window analysis of its input weights: waveform parameters dominate before 40% of ultimate strength, occurrence parameters dominate during 40-80%, and waveform parameters surge again near failure, with fractal dimension the most consistently important input. These weight shifts are interpreted as tracing three stages of fault evolution: nucleation of random microcracks, initiation of fault growth, and coalescence into a through-going fault.

Load-bearing premise

The load-bearing premise is that using the rock's peak strength and its strain at failure to set the time windows and axes of the training data does not secretly pass the network the answer it is supposed to predict; if that leakage is not harmless, the reported prediction accuracy is inflated.

Editorial extensions

If this is right

  • If the network's dynamic weighting reflects real mechanics, monitoring systems could flag the transition from distributed cracking to localised faulting simply by watching which class of acoustic feature dominates the model's attention.
  • The model's success at 5, 10, 20, and 40 MPa, spanning axial splitting and shear localisation, implies the feature set carries usable failure information across different failure mechanisms, not just one laboratory geometry.
  • The explained weight trajectories give a concrete target for physical validation: independent imaging or post-mortem inspection should show increased fracture clustering exactly where the network switches from waveform- to occurrence-dominated inputs.
  • If the scaling arguments hold, the same parameter set could be used to search for precursory changes in natural fault-zone monitoring, where peak delay, coda Q, b-value, and fractal dimension are already measurable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors do not test a causal version of their preprocessing, but doing so would settle how much of the reported accuracy is real forecasting: recompute all features in fixed elapsed-time windows, or with normalisation based only on data available up to each time step, and compare held-out error.
  • A natural extension is to turn the three-stage weight trajectory into a direct damage-stage classifier by training the network to output the stage label itself; the boundaries it learns could then be compared with X-ray tomography or thin-section observations.
  • The dependence of feature importance on confining pressure suggests that a field deployment would need pressure, depth, or stress state as an explicit input; the paper's four pressures are a start, not a complete scaling law.
  • Because fractal dimension requires localised event source locations, its high importance in the model may not transfer to field settings with sparse sensor arrays; a testable variant would retrain the network with fractal dimension estimated from a reduced or surface sensor geometry.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper trains a time-delay neural network (TDNN) on multi-parameter acoustic emission (AE) data from triaxial deformation experiments on Alzo granite at confining pressures of 5, 10, 20, and 40 MPa. Inputs are waveform-derived attributes (mean free path, coda Q, peak delay) and occurrence-based metrics (event rate, b-value, fractal dimension), while targets are strain and differential stress normalized to their values at failure. Hyperparameters are optimized with a genetic algorithm, and the trained network is used to predict normalized stress and strain as functions of the deformation state. The authors claim that the model predicts the timing and magnitude of stress drops across different pressures and failure mechanisms, and that the evolving importance of waveform versus occurrence parameters reveals distinct phases of fault nucleation, growth, and coalescence.

Significance. If the predictions were based on features that do not encode the target, the paper would offer a useful demonstration of how combined waveform and occurrence AE parameters can track laboratory fault development, and the explainability analysis would provide a physically interpretable ranking of parameters across deformation stages. However, the central claim is currently undermined by the construction of input features in windows defined by the ultimate compressive strength (UCS) and normalized to strain at failure, which embeds information about the target into the inputs. The validation set also contains only one sample per pressure, so the reported generalization across pressures is not established. The strengths of the paper—the integration of diverse AE attributes and the use of GA optimization—cannot be assessed fairly until these methodological issues are resolved.

major comments (4)
  1. [Training Data; Figure 1] The input features are not independent of the target. The text states that 'AE-derived parameters were smoothed using a moving window approach (20% UCS), with time-dependent variations normalised to strain at failure,' while the target variables are 'stress-strain mechanical data ... normalised against their values at sample failure (i.e., 100% UCS).' Because UCS and failure strain are only known at the end of the experiment, any model using these features cannot be considered predictive in a forward sense. The reported errors of ~5–10% UCS may simply reflect the fact that the inputs already encode the failure state. The authors must reanalyze the data with features computed using windows and normalizations that do not reference the failure point (e.g., fixed time windows or running medians without reference to UCS), and then re-evaluate whether the predictions remain accurate.
  2. [Neural Network Classification; Figures 2–3] The validation strategy is too weak to support the generalization claim. The model is trained on three experiments and validated on the fourth, with only one experiment per confining pressure (5, 10, 20, 40 MPa). With n=1 per condition, the apparent accuracy may reflect memorization of individual sample trends rather than a generalizable mapping from AE features to mechanical state. The authors should report the number of independent samples per pressure, perform repeated cross-validation (e.g., k-fold with multiple samples per fold), and provide confidence intervals or significance tests on the prediction errors. Without such analysis, the claim of prediction across varying pressures and failure mechanisms is not substantiated.
  3. [Neural Network Optimisation] The manuscript does not disclose the final TDNN architecture or the details of the genetic-algorithm optimization. The text mentions 'the hidden layer architecture' and 'hyperparameter architecture, the number of hidden layers, the number of neurons in each layer, and the time-delays for each layer,' but none of these quantities are reported for the models used in the results. The GA population size, number of generations, and the exact train/validation split used in the objective function are also omitted. This lack of detail prevents reproduction and makes it impossible to assess whether the reported performance arises from overfitting to the training experiments. The authors should provide the final architecture, the GA settings, and a description of the cross-validation procedure used during optimization.
  4. [Neural Network Parameter Sensitivity; Discussion] The parameter-importance analysis inherits the same target-leakage problem, since it is based on weights learned from features that are normalized to strain at failure and smoothed in UCS-defined windows. The physical interpretation of the importance curves (e.g., waveform parameters dominating before 30% UCS, occurrence parameters between 40–80% UCS) is therefore not trustworthy as evidence of a causal relationship. In addition, Figure 4 shows no uncertainty estimates; given that the model is trained over only four experiments, the reported importance variations may be statistically insignificant. The authors should either present a non-leaky version of the importance analysis with confidence intervals or temper the physical conclusions accordingly.
minor comments (7)
  1. [Introduction] The abbreviation 'TDNN' is used in the abstract and introduction without spelling out 'Time Delay Neural Network' at first occurrence; please define it explicitly at first use.
  2. [Acoustic Emission Attributes – Occurrence Derived] The event rate definition appears inconsistent: the text says 'inverse logarithm of inter-event timings' but the formula is 'e = ln(ΔT)', which is simply the logarithm of the time difference. Please clarify the sign and the intended relationship, and ensure the formula matches the description.
  3. [Figure 1 caption] The caption says the mechanical data are 'normalised against their values at sample failure (i.e., 100% UCS)', but UCS is a stress, whereas the y-axis of the strain plot is normalized strain. Please adjust the wording to distinguish the strain normalization from the stress normalization.
  4. [Materials and Methods; Neural Network Optimisation] The low-frequency and high-frequency band edges used for filtering the AE waveforms are never specified. Please provide the numerical values for the 'low and high frequency windows' so that the waveform-derived parameters can be reproduced.
  5. [Training Data] The 'strain shifts applied to align data across different experiments' are mentioned but not quantified. Please describe the magnitude and justification of these shifts, or state that they are applied on a per-experiment basis.
  6. [Data and materials availability] The data availability statement says that data and code are 'in a developmental format' and require collaboration; this falls short of standard journal expectations for sharing of data and code. Please provide a clear statement of what will be made available and under what conditions, or deposit the data and code in a public repository.
  7. [References] Several references are incomplete or lack page numbers (e.g., reference 29 is a book chapter with page range omitted). Please check the reference list against the journal's formatting requirements.

Circularity Check

1 steps flagged · score 7.0 of 10

Target-dependent preprocessing (UCS-based windows, normalization to strain at failure) makes the neural network's stress/strain 'predictions' partly circular.

  1. self definitional [Results, Training Data and Figure 1 caption; Methods, Acoustic Emission Attributes – Occurrence Derived]
    "AE-derived parameters were smoothed using a moving window approach (20% UCS), with time-dependent variations normalised to strain at failure. ... Stress-strain mechanical data recorded during the experiments are normalised against their values at sample failure (i.e., 100% UCS) and are used as the target classification in separate neural networks."

    The network's targets are stress and strain normalised to failure (100% UCS). The input features are not independent of these targets: every AE parameter is smoothed in moving windows whose width is defined as 20% UCS, and the time series is normalised to strain at failure. UCS and failure strain are exactly the quantities the network is trained to predict, so the input time base and window boundaries are functions of the target. At each normalized-strain position, the feature aggregates AE data over a stress-defined interval, embedding target information into the input construction. The reported 'prediction' of normalized stress/strain therefore reduces partly by construction to the normalization used to build the features.

full rationale

The main circularity is in the training-data construction. The paper's headline result is that a TDNN predicts 'the timing and magnitude of stress drops' from AE parameters. However, before training, every AE-derived input is 'smoothed using a moving window approach (20% UCS)' and 'normalised to strain at failure,' while the target is stress-strain 'normalised against their values at sample failure (i.e., 100% UCS).' UCS and failure strain are exactly the quantities the network is asked to predict. Using them to set the time base and window boundaries for the inputs means the feature set already contains the target's normalising constants and stress-dependent binning. In a real deployment one would not know UCS/failure strain a priori, so the claimed 'prediction' is not an independent forecast. The leave-one-pressure-out scheme only tests interpolation across pressures after all samples have been normalized by their own failure values, so it does not remove this leakage. The other candidate circularities—author self-citations for laboratory scattering methods and AE waveform picking—are not load-bearing for the central claim: they are methodological references to prior published work, not a uniqueness theorem or ansatz smuggled in to force the result. If the features were recomputed on absolute time/load without using UCS/failure strain, the TDNN result would be meaningful; as published, the core prediction partly reduces to the target-dependent preprocessing.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on a set of hand-chosen processing parameters (window size, frequency bands, convergence thresholds) and on domain assumptions about the physical meaning of the AE statistics. No new entities are introduced.

free parameters (5)
  • Moving window size for AE parameter extraction = 20% UCS
    Chosen by hand; all AE features are averaged over windows defined by the target stress-strain curve, introducing target information into the inputs.
  • TDNN architecture hyperparameters = GA-optimized, not reported
    The genetic algorithm evolves architectures; the final number of layers, neurons, and delays is never stated, so the model is under-specified.
  • GA and final training convergence criteria = 10% error / 30 s, then 1% error / 5 min
    Set through 'qualitative estimates of model accuracy during development' (Methods), i.e., hand-picked.
  • Low/high frequency band edges = Not specified
    The paper never defines the low and high frequency windows used for filtering the AE waveforms, leaving the waveform-derived features ambiguous.
  • Post-failure strain shifts = Experiment-dependent
    Applied to align data across experiments; this post-hoc shift uses target information to modify the input time series.
assumptions (4)
  • domain assumption AE wave energy transport follows the isotropic diffusion equation (Eq. 1).
    Used to derive mean free path and coda Q from waveform envelopes; if scattering is not isotropic, these features are not physically meaningful.
  • domain assumption Gutenberg-Richter b-value and box-counting fractal dimension computed from AE catalogs capture fault evolution.
    The analysis assumes these statistics are informative for stage identification; they are standard in the field but the interpretation is not independently verified.
  • domain assumption The three-stage fault evolution model (nucleation, growth, coalescence) applies to these experiments.
    Imported from Reches and Lockner (1994) and used to label the phases; the TDNN does not discover the phases, it is interpreted against them.
  • standard math Linear least squares inversion yields stable estimates of diffusion coefficient d and damping beta.
    Standard numerical inversion, but stability is assumed without reported error bars or convergence checks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalised Time-Series Analysis of Fault Mechanics Using Explainable AI." pith.science (2026). https://pith.science/paper/UUS67J3F

@misc{pith2026250521312,
  author       = {Pith},
  title        = {Pith review of: Generalised Time-Series Analysis of Fault Mechanics Using Explainable AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UUS67J3F}},
  note         = {Machine review of arXiv:2505.21312}
}
read the original abstract

Understanding how faults nucleate and grow is a critical problem in earthquake science and hazard assessment. This study examines fault development in Alzo granite under triaxial pressures ranging from 5 to 40 MPa by applying a Time Delay Neural Network (TDNN) to multi-parameter acoustic emission (AE) data. The TDNN integrates waveform-derived attributes, including peak delay and scattering attenuation, with occurrence-based metrics such as time distributions, Gutenberg-Richter b-values, and spatial fractal dimensions, to characterize the transition from distributed microcracking to localised faulting. Optimised via genetic algorithms, the TDNN dynamically weights these parameters, enabling accurate characterisation of fault growth stages. Our results delineate three distinct phases of fault evolution: nucleation of random microcracks indicated by changes in elastic wave scattering, initiation of fault growth reflected in evolving AE spatial and magnitude distributions, and fault coalescence marked by exponential increases in peak delay and b-value shifts. The model predicts the timing and magnitude of stress drops across varying pressures and failure mechanisms, from axial splitting to shear localisation, providing deeper insights into fault mechanics through explainable AI models.

Figures

Figures reproduced from arXiv: 2505.21312 by the authors.

Figure 2
Figure 2. For each confining pressure, the model was trained on data from three other tests and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Observed differential stress variations (black) normalised to stress at ultimate compressive strength (UCS, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    Reches, D

    Z. Reches, D. A. Lockner, Nucleation and growth of faults in brittle rocks. Journal of Geophysical Research: Solid Earth 99, 18159–18173 (1994)

    work page 1994

  2. [2]

    Reches, Mechanisms of slip nucleation during earthquakes

    Z. Reches, Mechanisms of slip nucleation during earthquakes. Earth and Planetary Science Letters 170, 475–486 (1999)

    work page 1999

  3. [3]

    B. D. Thompson, R. P. Young, D. A. Lockner, Fracture in Westerly Granite under AE Feedback and Constant Strain Rate Loading: Nucleation, Quasi-static Propagation, and the Transition to Unstable Fracture Propagation. pure and applied geophysics 163, 995–1019 (2006)

    work page 2006

  4. [4]

    Burlini, S

    L. Burlini, S. Vinciguerra, G. Di Toro, G. De Natale, P. Meredith, J.-P. Burg, Seismicity preceding volcanic eruptions: New experimental insights. Geology 35, 183–186 (2007)

    work page 2007

  5. [5]

    P. M. Benson, B. D. Thompson, P. G. Meredith, S. Vinciguerra, R. P. Young, Imaging slow failure in triaxially deformed Etna basalt using 3D acoustic‐emission location and X‐ray computed tomography. Geophysical Research Letters 34 (2007)

    work page 2007

  6. [6]

    X.-T. Feng, M. Seto, Fractal structure of the time distribution of microfracturing in rocks. Geophysical Journal International 136, 275–285 (1999)

    work page 1999

  7. [7]

    X. Lei, K. Masuda, O. Nishizawa, L. Jouniaux, L. Liu, W. Ma, T. Satoh, K. Kusunose, Detailed analysis of acoustic emission activity during catastrophic fracture of faults in rock. Journal of structural geology 26, 247–258 (2004)

    work page 2004

  8. [8]

    Lockner, J

    D. Lockner, J. D. Byerlee, V. Kuksenko, A. Ponomarev, A. Sidorin, Quasi-static fault growth and shear fracture energy in granite. Nature 350, 39–42 (1991)

    work page 1991

Show all 42 references
  1. [9]

    T. King, S. Vinciguerra, J. Burgess, P. Benson, L. De Siena, Source mechanisms of laboratory earthquakes during fault nucleation and formation. Journal of Geophysical Research: Solid Earth 126, e2020JB021059 (2021)

    2021

  2. [10]

    Renard, N

    F. Renard, N. Kandula, J. McBeck, B. Cordonnier, Creep burst coincident with faulting in marble observed in 4‐D synchrotron X‐ray imaging triaxial compression experiments. Journal of Geophysical Research: Solid Earth 125, e2020JB020354 (2020)

    2020

  3. [11]

    F. X. Passelègue, A. Schubnel, S. Nielsen, H. S. Bhat, D. Deldicque, R. Madariaga, Dynamic rupture processes inferred from laboratory microearthquakes. Journal of Geophysical Research: Solid Earth 121, 4343–4365 (2016)

    2016

  4. [12]

    Gabrielli, A

    S. Gabrielli, A. Akinci, L. De Siena, E. Del Pezzo, M. Buttinelli, F. E. Maesano, R. Maffucci, Scattering attenuation images of the control of thrusts and fluid overpressure on the 2016–2017 Central Italy seismic sequence. Geophysical Research Letters 50, e2023GL103132 (2023)

    2023

  5. [13]

    Nardoni, P

    C. Nardoni, P. Persaud, Evidence for faulting and fluid-driven earthquake processes from seismic attenuation variations beneath metropolitan Los Angeles. Scientific Reports 14, 17595 (2024)

    2024

  6. [14]

    Sketsiou, F

    P. Sketsiou, F. Napolitano, A. Zenonos, L. De Siena, New insights into seismic absorption imaging. Physics of the Earth and Planetary Interiors 298, 106337 (2020)

    2020

  7. [15]

    T. King, L. De Siena, P. Benson, S. Vinciguerra, Mapping faults in the laboratory with seismic scattering 1: the laboratory perspective. Geophysical Journal International 232, 1590–1599 (2023)

    2023

  8. [16]

    T. King, L. De Siena, Y. Zhang, N. Nakata, P. Benson, S. Vinciguerra, Mapping faults in the laboratory with seismic scattering 2: The modelling perspective. Geophysical Journal International 234, 1024–1031 (2023)

    2023

  9. [17]

    X. Lei, S. Ma, Laboratory acoustic emission study for earthquake generation process. Earthquake Science 27, 627–646 (2014)

    2014

  10. [18]

    Karimpouli, G

    S. Karimpouli, G. Kwiatek, P. Martínez-Garzón, G. Dresen, M. Bohnhoff, Unsupervised clustering of catalogue-driven features for characterizing temporal evolution of labquake stress. Geophysical Journal International 237, 755–771 (2024)

    2024

  11. [19]

    Rouet‐Leduc, C

    B. Rouet‐Leduc, C. Hulbert, D. C. Bolton, C. X. Ren, J. Riviere, C. Marone, R. A. Guyer, P. A. Johnson, Estimating fault friction from seismic signals in the laboratory. Geophysical Research Letters 45, 1321–1329 (2018)

    2018

  12. [20]

    S. M. Mousavi, G. C. Beroza, Machine learning in earthquake seismology. Annual Review of Earth and Planetary Sciences 51, 105–129 (2023)

    2023

  13. [21]

    T. King, P. Benson, L. De Siena, S. Vinciguerra, Acoustic emission waveform picking with time delay neural networks during rock deformation laboratory experiments. Seismological Research Letters 92, 923–932 (2021)

    2021

  14. [22]

    Lubbers, D

    N. Lubbers, D. C. Bolton, J. Mohd‐Yusof, C. Marone, K. Barros, P. A. Johnson, Earthquake catalog‐based machine learning identification of laboratory fault states and the effects of magnitude of completeness. Geophysical Research Letters 45, 13,269-13,276 (2018)

    2018

  15. [23]

    D. C. Bolton, P. Shokouhi, B. Rouet‐Leduc, C. Hulbert, J. Rivière, C. Marone, P. A. Johnson, Characterizing acoustic signals and searching for precursors during the laboratory seismic cycle using unsupervised machine learning. Seismological Research Letters 90, 1088–1098 (2019)

    2019

  16. [24]

    Jasperson, D

    H. Jasperson, D. C. Bolton, P. Johnson, R. Guyer, C. Marone, M. V. de Hoop, Attention network forecasts time‐to‐failure in laboratory shear experiments. Journal of Geophysical Research: Solid Earth 126, e2021JB022195 (2021)

    2021

  17. [25]

    Laurenti, E

    L. Laurenti, E. Tinti, F. Galasso, L. Franco, C. Marone, Deep learning for laboratory earthquake prediction and autoregressive forecasting of fault zone stress. Earth and Planetary Science Letters 598, 117825 (2022)

    2022

  18. [26]

    Karimpouli, D

    S. Karimpouli, D. Caus, H. Grover, P. Martínez-Garzón, M. Bohnhoff, G. C. Beroza, G. Dresen, T. Goebel, T. Weigel, G. Kwiatek, Explainable machine learning for labquake prediction using catalog-driven features. Earth and Planetary Science Letters 622, 118383 (2023)

    2023

  19. [27]

    Shokouhi, V

    P. Shokouhi, V. Girkar, J. Rivière, S. Shreedharan, C. Marone, C. L. Giles, D. Kifer, Deep Learning Can Predict Laboratory Quakes From Active Source Seismic Data. Geophysical Research Letters 48, e2021GL093187 (2021)

    2021

  20. [28]

    Shreedharan, D

    S. Shreedharan, D. C. Bolton, J. Rivière, C. Marone, Machine Learning Predicts the Timing and Shear Stress Evolution of Lab Earthquakes Using Active Seismic Monitoring of Fault Zone Processes. Journal of Geophysical Research: Solid Earth 126, e2020JB021588 (2021)

    2021

  21. [29]

    Phoneme Recognition Using Time-Delay Neural Networks

    A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, K. J. Lang, “Phoneme Recognition Using Time-Delay Neural Networks” in Backpropagation: Theory, Architectures, and Applications, Y. Chauvin, D. E. Rumelhart, Eds. (Psychology Press, 1995; https://www.taylorfrancis.com/chapters/edit...

    1995 doi

  22. [30]

    JHU ASpIRE system: Robust LVCSR with TDNNS, iVector adaptation and RNN-LMS

    V. Peddinti, G. Chen, V. Manohar, T. Ko, D. Povey, S. Khudanpur, “JHU ASpIRE system: Robust LVCSR with TDNNS, iVector adaptation and RNN-LMS” in 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU) (2015), pp. 539– 546

    2015

  23. [31]

    Continuous time delay neural networks for detection of temporal patterns in signals

    R. Derakhshani, S. A. C. Schuckers, “Continuous time delay neural networks for detection of temporal patterns in signals” in 2004 IEEE International Joint Conference on Neural Networks (IEEE Cat. No.04CH37541) (2004)vol. 4, pp. 2723–2728 vol.4

    2004

  24. [32]

    Takahashi, H

    T. Takahashi, H. Sato, T. Nishimura, K. Obara, Tomographic inversion of the peak delay times to reveal random velocity fluctuations in the lithosphere: Method and application to northeastern Japan. Geophysical Journal International 178, 1437–1455 (2009)

    2009

  25. [33]

    Wegler, Diffusion of seismic waves in a thick layer: Theory and application to Vesuvius volcano

    U. Wegler, Diffusion of seismic waves in a thick layer: Theory and application to Vesuvius volcano. Journal of Geophysical Research: Solid Earth 109 (2004)

    2004

  26. [34]

    L. Ding, Y. Zhao, Y. Pan, Y. Fan, Investigation on acoustic emission characteristics of fault stick-slip under different lateral pressures. Scientific Reports 14, 6718 (2024)

    2024

  27. [35]

    Vinciguerra, S

    S. Vinciguerra, S. Gresta, M. S. Barbano, G. Distefano, The two behaviours of Mt. Etna Volcano before and after a large intrusive episode: Evidences from b value and fractal dimension of seismicity. Geophysical Research Letters 28, 2257–2260 (2001)

    2001

  28. [36]

    - White granites used in lombard architecture

    R. Bugini, A. Pavese, S. Borroni, L. Folli, “- White granites used in lombard architecture” in Proceedings of the 9th International Congress on Deterioration and Conservation of Stone, V. Fassina, Ed. (Elsevier Science B.V., Amsterdam, 2000; https://www.sciencedirect.com/scien...

    2000

  29. [37]

    C. E. Harnett, P. M. Benson, P. Rowley, M. Fazio, Fracture and damage localization in volcanic edifice rocks from El Hierro, Stromboli and Tenerife. Scientific reports 8, 1942 (2018)

    2018

  30. [38]

    Voight, A Relation to Describe Rate-Dependent Material Failure

    B. Voight, A Relation to Describe Rate-Dependent Material Failure. Science 243, 200– 203 (1989)

    1989

  31. [39]

    C. R. J. Kilburn, Multiscale fracturing as a key to forecasting volcanic eruptions. Journal of Volcanology and Geothermal Research 125, 271–289 (2003)

    2003

  32. [40]

    G. M. Atkinson, D. W. Eaton, N. Igonin, Developments in understanding seismicity triggered by hydraulic fracturing. Nature Reviews Earth & Environment 1, 264–277 (2020)

    2020

  33. [41]

    X. Kong, E. Wang, S. Li, H. Lin, P. Xiao, K. Zhang, Fractals and chaos characteristics of acoustic emission energy about gas-bearing coal during loaded failure. Fractals 27, 1950072 (2019)

    2019

  34. [42]

    Dogru, E

    A. Dogru, E. Gorgun, B. Aktug, H. Ozener, Seismic hazard assessment of the central North Anatolian Fault (Turkey) from GPS-derived strain rates and b-values. Geomatics, Natural Hazards and Risk 9, 356–367 (2018). Acknowledgments The authors would like to thank numerous colleag...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.