Deep learning for fast simulation of seismic waves in complex media

Moseley, Ben; Nissen-Meyer, Tarje; Markham, Andrew

doi:https://doi.org/10.5194/se-11-1527-2020

Articles | Volume 11, issue 4

https://doi.org/10.5194/se-11-1527-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/se-11-1527-2020

© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 11, issue 4

Research article

|

24 Aug 2020

Research article |

| 24 Aug 2020

Deep learning for fast simulation of seismic waves in complex media

Ben Moseley, Tarje Nissen-Meyer, and Andrew Markham

Download

Final revised paper (published on 24 Aug 2020)
Preprint (discussion started on 13 Nov 2019)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of manuscript by Moseley et al.', Andrew Valentine, 22 Dec 2019
- AC1: 'Response to referee comments', Ben Moseley, 25 Jan 2020
RC2: 'Review of Moseley et al., 2019', Andrew Curtis, 24 Dec 2019
- AC1: 'Response to referee comments', Ben Moseley, 25 Jan 2020

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Ben Moseley on behalf of the Authors (20 Apr 2020) Author's response Manuscript

ED: Referee Nomination & Report Request started (01 May 2020) by Caroline Beghein

RR by Andrew Curtis (11 May 2020)

RR by Andrew Valentine (03 Jun 2020)

Suggestions for revision or reasons for rejection

This manuscript has evolved significantly since its original submission, and I thank the authors for their careful consideration of the reviews. A few comments:
- The authors frame their manuscript as 'a manifesto or roadmap for deep learning'. I think this is a good line to take, and their work certainly makes a convincing case for further exploration. However, it should be noted that the manuscript really only explores one or two possible ways of framing seismic problems using deep learning. On occasion the authors' language creeps beyond this, towards a claim that they have comprehensively characterised the space of possibilities.
- I think the comparisons between the wavenet-derived simulations and 2D ray tracing are very interesting. If I understand Fig. A1 correctly, the authors' approach consistently out-performs the 'physical' method in terms of accuracy/fidelity to the training data. This could perhaps be discussed in more detail in the main text. Looking at Fig. 5, it appears that the ray-tracing introduces a slight phase shift relative to the reference waveforms. Is this the main source of discrepancy between the two? What do the worst-performing wavenet simulations look like? Does wavenet always 'more-or-less' work, or does it sometimes fail egregiously?
- To what extent have the authors explored different deep learning formulations? Does using a smaller NN cause gradual deterioration or total failure of the approach?
- The section on 'inverse wavenet' (using deep learning to map waveforms into models) has been scaled back significantly as a result of the earlier reviews. I wonder if it might be appropriate to move all discussion of this (i.e. formulation, training, results, discussion) into a single place within the manuscript, presumably within Section 4. I think this might better reflect the speculative tone of this part.

Finally, as a more general perspective: I would argue that the utility (or otherwise) of 'learned' approaches is highly context-dependent. In general, machine-learning--based methods offer computational costs that scale very differently from 'conventional' approaches, but with less favourable (or at least, less clearly-understood) accuracy guarantees. I think this probably means that further development of 'learned' methods needs to be targeted towards specific identified applications: the cost/benefit analysis for (say) applications in earthquake early warning may be very different from those in (say) global tomography. Optimising and validating any learned strategy ought to take this into account, and I am not sure that a one-size-fits-all approach is realistic.

Hide

ED: Publish subject to minor revisions (review by editor) (08 Jun 2020) by Caroline Beghein

Dear authors,

Your manuscript was sent back to the reviewers who both agreed you have done a good job of addressing their comments. I will be ready to accept your paper for publication after a few minor changes suggested by the reviewers if at all possible. See below:

- From reviewer 1:
When discussing the extension to 3D, the authors state: "However, this approach is likely to be practically challenging because increasing the dimensionality would increase the number of weights and likely the training time." I think they should mention a little more explicitly/clearly that there are three compounding reasons that costs increase when moving to 3D, of which I think the increased number of weights is probably the least of our worries...:
1. the number of weights in the network increases significantly.
2. the cost of forward modelling the training examples increases hugely.
3. the "curse of dimensionality" needs to be mentioned explicitly as this is what implies that we require exponentially many more samples and hence modelling runs in order to explore parameter space.

-From reviewer 2:
I think the comparisons between the wavenet-derived simulations and 2D ray tracing are very interesting. If I understand Fig. A1 correctly, the authors' approach consistently out-performs the 'physical' method in terms of accuracy/fidelity to the training data. This could perhaps be discussed in more detail in the main text. Looking at Fig. 5, it appears that the ray-tracing introduces a slight phase shift relative to the reference waveforms. Is this the main source of discrepancy between the two? What do the worst-performing wavenet simulations look like? Does wavenet always 'more-or-less' work, or does it sometimes fail egregiously?
- To what extent have the authors explored different deep learning formulations? Does using a smaller NN cause gradual deterioration or total failure of the approach?
- The section on 'inverse wavenet' (using deep learning to map waveforms into models) has been scaled back significantly as a result of the earlier reviews. I wonder if it might be appropriate to move all discussion of this (i.e. formulation, training, results, discussion) into a single place within the manuscript, presumably within Section 4. I think this might better reflect the speculative tone of this part.

Thank you.

Caroline Beghein

Hide

AR by Ben Moseley on behalf of the Authors (21 Jun 2020) Author's response Manuscript

ED: Publish as is (26 Jun 2020) by Caroline Beghein

ED: Publish as is (30 Jun 2020) by CharLotte Krawczyk (Executive editor)

AR by Ben Moseley on behalf of the Authors (04 Jul 2020) Author's response Manuscript

Short summary

Simulations of seismic waves are very important; they allow us to understand how earthquakes spread and how the interior of the Earth is structured. However, whilst powerful, existing simulation methods usually require a large amount of computational power and time to run. In this research, we use modern machine learning techniques to accelerate these calculations inside complex models of the Earth.