compgeolab / eql-gradient-boosted Goto Github PK

Pin Harmonica to v0.2.0 and Verde to v1.6.0 after they are released.

Information

• Journal name: Geophysical Journal International
• Journal guidelines for authors: https://academic.oup.com/gji/pages/General_Instructions
• Supplementary material DOI: https://doi.org/10.6084/m9.figshare.13604360

This is the standard EarthArXiv disclaimer for a preprint:

\noindent
\textbf{Disclaimer:}
This is a non-peer reviewed preprint submitted to EarthArXiv.
It is currently under review at \textit{insert-journal}.

Checklist

• Reserve a DOI on figshare or Zenodo for the supplementary material (zip archive of this repository). Paste the DOI at the top of this issue.
• Add the supplementary DOI to the manuscript
• Make a cover letter for the editor. Include a short summary, suggested reviewers, and contributions by student authors
• Submit the paper to the Journal
• Submit the preprint to EarthArXiv
• Add the preprint DOI to the README.md
• Make a release on GitHub
• Upload the .zip version of the repository to figshare or Zenodo and publish it
• Add the preprint to the CompGeoLab (and your personal) website
• Tell people about your preprint on social media (include a figure and maybe a thread with the main results)

@santisoler we forgot a very important thing! Adding some random noise to the data. I suspect this will drive the scores down a bit but I don't think it will change the results much.

The matplotlib.rc file was created so several notebooks can use the same configuration for creating figures for the manuscript. Since we centrilised the figures creation on a single notebook, there's no need to maintain this file, we could simply change matplotlib's configuration on the notebook itself.

We decided to keep only with the Australia gravity data for a real world application.

A checklist of things to do once the paper is accepted:

• Update the preprint disclaimer to include the version of reccrd
• Update the README with the version of record link
• Update the figshare archive
• Tag a new version as accepted
• Update the preprint on EarthArXiv
• Sign and submit the publisher agreement
• Double-check the source files and update any that need changing on the journal system

From this recent paper: https://www.pinga-lab.org/papers/convolutional-eql-grav.html

Based on the results of testing the performance of the different source distributions, we can conclude that:

• Block-averaging the sources reduces the computational load, while keeping good accuracy. It's computationally cheaper than using source below data or grid_sources
• Changing the depth type doesn't introduce any considerable improvement on the results.
• Using variable depths might be more expensive when searching the best set of parameters through CV.
• Using relative depth or constant depth reduces the number of parameters to the minimum (depth and damping), while generating the same level of accuracy.

So, would be better to stick with relative depth on any further application of the gridders, like the ones we do with EQLHarmonicBoost.

@birocoles and @valcris suggested plotting normalize singular values for the Jacobians in each source distribution type. This will help give a more theoretical basis for our conclusions on which layout is better.

After working with jupytext for a long time, we haven't experienced much of a gain.
The process of keeping both the .py files and the .ipynb notebooks updated had caused some problems and the benefit of review can be achieved through other tools with less hassle.

We don't mention this at all but it would be good to round out the introduction. Need to first do some reading on collocation 🙂

See what GJI says about that and make sure we comply.

I need to finish my review of the methods section.

After submission, we could post a preprint to EarthArXiv if desired. An advantage of doing this is that we could start coding the gradient boosting into Harmonica and cite the preprint while we wait for the paper reviews.

If we do this, we need to format the paper for EarthArXiv by removing the GJI style and putting a disclaimer at the top saying that it's a preprint that hasn't been peer-reviewed.

@santisoler what do you think?

Following #60, we would need to move from R2 score to RMS also on cross validation of EQL gridders, which is only used on real world cases (Australia gravity for example, on #61).

To be able to use a different score on cross validation we would need to use the latest version of Verde that introduces a new scoring argument to verde.cross_val_score.

I think we can safely pass sklearn.metrics.mean_squared_error, although new we would set the best predictor as the one that achieves the minimum score. Also remember to compute its square root before appending it to the scores list.

Jupytext is awesome to actually get a diff of the notebooks. But it also kind of sucks when I just want to see your results without having to run the code here (for example, it might take a long time). One alternative would be keep both the .py and .ipynb files in the history. We can review in PRs using the .py but be able to quickly check the outputs in the notebook. It's a bit wasteful but it's not like we'll be commiting to this repo for years.

What do you think?

Add some experiments showing how the gradient boosting algorithm behaves when changing the window size on synthetic data.
Topics of interest would be:

• How accurate predictions are for small windows? Is there a threshold where predictions are garbage?
• How window size affects computational load?

The last one might be tricky because increasing the window reduces the number of iterations, but increases the size of the least square problem. Might be a trade-off between these two.

Just to register some ideas:

1. Reducing the number of parameters in equivalent source processing

From the GJI instructions:

Author contribution statement: GJI encourages authors to include an author contribution statement, stating for example: who analysed the results, who processed the data, who wrote the paper. This should be included as a narrative statement in the acknowledgements section of the paper.

Questions from the chat session:

Peter Lelievre, Memorial University (16:19) Hello Santiago. I'm happy to see someone investigating these important practical questions. I have two sets of comments/questions. First, regarding the x's in the table on slide 20 of your presentation: combining the grid source layout with relative depth positioning should be possible: just interpolate onto the topography elevations and then push downwards by some constant amount, no? I feel like combining the grid source layout with variable depth positioning should also be possible: can you use the same strategy you show on page 18 and calculate the mean distance of the gridded points to their nearest neighbours in the original measurement locations?

Hans-Juergen Goetze CAU Kiel (audience) (16:17) Santiago: Could 3D kriging possibly be an acceptable method?

Paolo Mancinelli UniChieti (16:20) Have you tested it on real case datasets?

Apply the correct style files.

After submission, we should make this repository public so that reviewers can see the code.