Deep Learning for Seismology

Over a relatively short span of time, machine learning has become an indispensible tool in seismology and geophysics. Benefitting from the vast data sets available to the community, machine learning as a data-driven approach excels in performing non-linear tasks such as automated earthquake detection, phase picking, and signal processing. Very often these tasks are solved with supervised learning methods, which require a ground truth to train machine learning models.

Instead of focusing on big models trained on big data, can we instead make new discoveries and advances in a self-supervised way?

Recently, I started to explore a number of avenues that address this question. To help me navigate through the rapidly evolving landscape of machine learning, I take inspiration from recent advances in, for instance, signal and image processing, astrophysics, and fluid mechanics. With those computational concepts in the back of my mind, I revisit seismological problems such as seismic array beamforming, earthquake detection, and seismic source characterisation.

Key publication van den Ende et al. (2021), A Self-Supervised Deep Learning Approach for Blind Denoising and Waveform Coherence Enhancement in Distributed Acoustic Sensing Data, IEEE TNNLS, doi:10.1109/TNNLS.2021.3132832 [preprint]

Distributed Acoustic Sensing

Being the backbone of our digital infrastructure, fibre-optic cables can be found everywhere on Earth, both on land and in the oceans. When you send a laser pulse through an optical fibre, tiny imperfections in the fibre cause scattering of the light that can be measured as "optical echoes". By analysing these echoes, one can infer stretching and shrinking of the fibre at different locations along the cable, a technique called Distributed Acoustic Sensing (DAS).

We can use DAS to record various types of vibrations along the cable, including earthquakes, cars (when the cable is on land), and ocean waves and boats when the cable is underwater. This opens up a plethora of possibilities to use existing fibre-optic cables (which we use for our everyday internet) as sensitive seismometers.

Key publication van den Ende & Ampuero (2021), Evaluating Seismic Beamforming Capabilities of Distributed Acoustic Sensing Arrays, Solid Earth, doi:10.5194/se-12-915-2021 [Open Access]

Earthquake Source Mechanics

Earthquakes (which can potentially grow to hundreds of kilometres in size) are the catastrophic result of processes that operate at the scale of micrometres (and even smaller). Even though we cannot directly resolve these processes with seismological methods, efforts have been made to explore the connection between the micro-scale and kilometre-scale through numerical modelling.

Together with other collaborators, I have developed a quasi-dynamic earthquake cycle code: QDYN. With this seismic cycle simulator, we can simulate numerous earthquake cycles and study the relation between local fault properties, such as fault zone width and rock solubility, and large scale emergent behaviour, such as slow earthquakes and run-away ruptures. Similarly, this code can be used to tackle inverse problems, extracting fault zone properties that explain natural observations of earthquake ruptures. The QDYN code is open-source available from GitHub.

Key publication van den Ende et al. (2018), A comparison between rate-and-state friction and microphysical models, based on numerical simulations of fault slip, Tectonophysics 733(9), doi:10.1016/j.tecto.2017.11.040 [Open Access]

• van den Ende et al. (2021), A Self-Supervised Deep Learning Approach for Blind Denoising and Waveform Coherence Enhancement in Distributed Acoustic Sensing Data, IEEE TNNLS, doi:10.1109/TNNLS.2021.3132832 [preprint]
• van den Ende & Ampuero (2021), Evaluating seismic beamforming capabilities of Distributed Acoustic Sensing Arrays, Solid Earth, doi:10.5194/se-12-915-2021 [Open Access]
• van den Ende et al. (2020), Extracting microphysical fault friction parameters from laboratory- and field injection experiments, Solid Earth, doi:10.5194/se-11-2245-2020 [Open Access]
• van den Ende & Ampuero (2020), Automated Seismic Source Characterisation Using Deep Graph Neural Networks, Geophysical Research Letters, doi:10.1029/2020GL088690 [preprint]
• van den Ende et al. (2020), Rheological Transitions Facilitate Fault‐Spanning Ruptures on Seismically Active and Creeping Faults, Journal of Geophysical Research: Solid Earth 125(8), doi:10.1029/2019JB019328 [preprint]
• van den Ende & Ampuero (2020), On the Statistical Significance of Foreshock Sequences in Southern California, Geophysical Research Letters 47(3), doi:10.1029/2019GL086224 [preprint]
• van den Ende & Niemeijer (2019), An investigation into the role of time-dependent cohesion in interseismic fault restrengthening, Scientific Reports 9, doi:10.1038/s41598-019-46241-5 [Open Access]
• van den Ende et al. (2019), Influence of grain boundary structural evolution on pressure solution creep rates, Journal of Geophysical Research: Solid Earth 124(10), doi:10.1029/2019JB017500 [preprint]
• van den Ende & Niemeijer (2018), Time‐dependent compaction as a mechanism for regular stick‐slips, Geophysical Research Letters 45(12), doi:10.1029/2018GL078103 [Open Access]
• van den Ende et al. (2018), A comparison between rate-and-state friction and microphysical models, based on numerical simulations of fault slip, Tectonophysics 733(9), doi:10.1016/j.tecto.2017.11.040 [Open Access]