Outcomes meeting 17/10 - Poster

[amartinhuertas]

• Address @sghelichkhani 's comments (https://github.com/GiteonCaulfied/COMP4560_stokes_ml_project/issues/11)
• Move the current bullet points in the introduction-Motivation as paragraphs to the project report. Reduce the amount of text in this poster's section. I have the impression that part of the current sentences in these paragraphs can be refactored in the subsection below entitled "Mantle Convection Problem".
• Use geoid with lower case g except those standard points where upper case G is needed.
• Figure 2.
  • Substitute equations by the one provided by Sia's.
  • Replace Parameters box by a picture of the initial temperature field.
  • Label the box on the left with "initial temperature", and the box on the right with "temperature time evolution predicted by numerical solver" [For simplicity, perhaps we can skip the illustration about the random generation of initial conditions by sampling a probability distribution, or we can have it in the data set section, if we believe it's going to be necessary, helpful]
• Research aims of the Mantle Convection Problem @sghelichkhani
• "Methods (Continue)". "Two ways ..."->"Two possible ways ..."
• Figure 7. Try to find a more serious reference to the figure on the left. If not possible, replace in the caption. "Source of the left figure: StackOverflow".
• Rewrite "The maximum value of S that is not significantly affecting accuracy should be 4 since the sum of loss and the sum of POD (Proper Orthogonal Decomposition) difference tend to have a rapid increase when S is increased from 4 to 8." by "The growth of the loss function (accuracy degradation) with S is remarkably moderate, with only a few outliers deviating moderately from the median". If we can replace "remarkably moderate" by an estimated "rate of growth" experimentally determined, much better. @rhyshawkins
• Future works (Mantle convection).
  • To further investigate the outliers in Figure 11 to see whether a common pattern can be identified.
  • Consider more complex physics (e.g., non-linear constitutive laws) in the Mantle convection model.
  • To investigate to which extent the current surrogate forward models are accurate enough to end up with sensible, application experts-certified solutions to the inverse Mantle Convection problem.

[rhyshawkins]

Just pushed an extra demo of plotting using cartopy to SphericalHarmonics directory. You need to do a "pip install cartopy" to run.

[GiteonCaulfied]

Hi @amartinhuertas @rhyshawkins,

I've pushed an updated poster to the repo. The Conclusion section is now complete and right now the only missing part is the Research Aims of the Mantle Convection Problem.

• Use geoid with lower case g except those standard points where upper case G is needed.

• Figure 2.

  • Substitute equations by the one provided by Sia's.
  • Replace Parameters box by a picture of the initial temperature field.
  • Label the box on the left with "initial temperature", and the box on the right with "temperature time evolution predicted by numerical solver" [For simplicity, perhaps we can skip the illustration about the random generation of initial conditions by sampling a probability distribution, or we can have it in the data set section, if we believe it's going to be necessary, helpful]
• "Methods (Continue)". "Two ways ..."->"Two possible ways ..."
• Rewrite "The maximum value of S that is not significantly affecting accuracy should be 4 since the sum of loss and the sum of POD (Proper Orthogonal Decomposition) difference tend to have a rapid increase when S is increased from 4 to 8." by "The growth of the loss function (accuracy degradation) with S is remarkably moderate, with only a few outliers deviating moderately from the median". If we can replace "remarkably moderate" by an estimated "rate of growth" experimentally determined, much better. @rhyshawkins

• Future works (Mantle convection).

  • To further investigate the outliers in Figure 11 to see whether a common pattern can be identified.
  • Consider more complex physics (e.g., non-linear constitutive laws) in the Mantle convection model.
  • To investigate to which extent the current surrogate forward models are accurate enough to end up with sensible, application experts-certified solutions to the inverse Mantle Convection problem.

All of these have been fixed and updated. The figures are now replaced with the ones generated with @rhyshawkins 's latest code.

• Figure 7. Try to find a more serious reference to the figure on the left. If not possible, replace in the caption. "Source of the left figure: StackOverflow".

I am still trying to find a more serious reference, so right now I just replace in the caption: "Source of the left figure: StackOverflow"

• Move the current bullet points in the introduction-Motivation as paragraphs to the project report. Reduce the amount of text in this poster's section. I have the impression that part of the current sentences in these paragraphs can be refactored in the subsection below entitled "Mantle Convection Problem".

The motivation has been reduced to smaller size, as for other text I'll try my best to cut down the words I have there. For the "Mantle Convection Problem" subsection in the Introduction, I only add a high level description from @sghelichkhani 's motivation since the figure itself is very intuitive and right now I can't think of other things to add there.

• Research aims of the Mantle Convection Problem @sghelichkhani

Unfortunately, I am having trouble remembering the research aims of the Mantle Convection Problem discussed in the meeting even though @sghelichkhani mentioned in his issue #11 that "aim should be put as we discussed in here". It would be really nice if @sghelichkhani could add a copy of the research aims in this issue as @rhyshawkins did earlier.

[amartinhuertas]

Some additional comments:

• Fig. 1: Use "3D" instead of "Three-dimensional"
• Fig. 1. Title. What do you mean by "gravitational equipotential figure". Do you want to say here "field" instead of "figure"?
• Fig 2. Right-most picture title. Use "Temperature" instead of "temperature".
• Fig. 10. "First method" and "Second method" are not intention-revealing names. Can we try to find a meaningful name for these methods?

[amartinhuertas]

Another concern I have.

There seems to be quite a lot of white space among blocks, and the font size also seems to be quite small.

Have you seen an example poster to see whether this is what is actually expected?

[amartinhuertas]

Fig 8 and Fig 9: are we really visualizing the coefficients or the geoid field itself? Note that the field is an expansion using the coefficients in terms of the spherical harmonics.

[GiteonCaulfied]

Hi @amartinhuertas

The newest poster has been updated to repo.

• Fig. 1: Use "3D" instead of "Three-dimensional"
• Fig. 1. Title. What do you mean by "gravitational equipotential figure". Do you want to say here "field" instead of "figure"?
• Fig 2. Right-most picture title. Use "Temperature" instead of "temperature".

Fig 8 and Fig 9: are we really visualizing the coefficients or the geoid field itself? Note that the field is an expansion using the coefficients in terms of the spherical harmonics.

All these have been fixed.

• Fig. 10. "First method" and "Second method" are not intention-revealing names. Can we try to find a meaningful name for these methods?

I've now renamed the "First method" and "Second method" as "One-for-All" and "One-for-One", which I believe is kinda intention-revealing.

Another concern I have.

There seems to be quite a lot of white space among blocks, and the font size also seems to be quite small.

Have you seen an example poster to see whether this is what is actually expected?

I checked the sample poster and I have to say that poster is way more appealing than mine (even though Michael said in the Wattle discussion that it was not necessarily typical), I will try to find a way to make mine look better as well.

Additionally, to reduce the amount of text, the following things have been removed from the poster as I believe they are not that important for an appealing poster and not listed in the Poster specification on Wattle (I have a copy of them so that I can always add them back if you think I deleted something important):

• Literature
• Future Work
• Neural network background in the Introduction section
• Two conclusions that I think it's kinda redundant
• Some wordy captions

Also, the text font has been increased from 14pt to 17pt.

[GiteonCaulfied]

Hi @amartinhuertas @rhyshawkins,

The plot showing the growth of average errors with respect to S using LogLog has been added to the report. The growth of error is less than O(S) and is approximately O(0.77S)

I also made a few adjustment according to the comment I received from @rhyshawkins this morning, including:

• Putting back the Neural network background in the Introduction section
• Putting back the Literature section (now it's named as Reference)
• Added a short paragraph for the Research Aim of Mantle Convection Problem
• other minor fixes

[amartinhuertas]

Ok.

Another concern I have. We have a list of references at the end. But these are not referenced in the poster. Perhaps we could consider to have a section of related work? This would help to contextualize the research.

[GiteonCaulfied]

Hi @amartinhuertas,

I put back the references because @rhyshawkins mentioned in his comment that "A poster will often have up to approx 5 references". I tried to spare some space for the related work section, however, the poster is already pretty full and I can't seem to clear more space for a complete related work section :(

Also, I added the names of supervisors in the upper right corner.

UPDATE: I'll remove the current references and put other references there, such as the geoid background journal that @rhyshawkins mentioned earlier + some references about NNs.

UPDATE of UPDATE: the references have now been updated.

[amartinhuertas]

I'll remove the current references

TBH, I don't think removing highly related work references is a good idea.

I would add a sentence ... along the lines ... (please adapt) "Previous works have leveraged FNNs/CNNs to solve the inverse Geoid problem directly [MSc thesis] and FNNs/LSTMs to approximate the forward Mantle convention problem [Mantle convection in Mars refs]. However, our approach is different to these works, as we aim to combine the adjoint method with ML techniques, and besides, at producing whole time-series from the initial temperature field, instead of a single time step forward out a sequence of previous time-steps given a set of model parameter values."

[GiteonCaulfied]

Hi @amartinhuertas,

A short paragraph about the related works has been added to the start of the Method section in the Poster. I also put back two references I deleted, thanks for the suggestions!

[amartinhuertas]

A short paragraph about the related works has been added to the start of the Method section in the Poster. I also put back two references I deleted, thanks for the suggestions!

I removed one of the references (my reference, not needed), and made some changes in the related work paragraphs. Please review.

[GiteonCaulfied]

Hi @amartinhuertas,

I have reviewed all the changes and they seem great, thanks for the effort!

[amartinhuertas]

I have reviewed all the changes and they seem great, thanks for the effort!

FYI ... I just pushed some minor touches.

In regards to this:

In contrast to these works, we use 1D spherically symmetric viscosity model to solve the geoid problem.

are you 100% sure this is is accurate?

[GiteonCaulfied]

Hi @amartinhuertas,

In contrast to these works, we use 1D spherically symmetric viscosity model to solve the geoid problem.

are you 100% sure this is is accurate?

I take another look at the master thesis, and I think one of the main difference between their work and ours is that we use FNN instead of CNN. Therefore, I've modified the above sentence as "we use FNN on 1D..."

[amartinhuertas]

I take another look at the master thesis, and I think one of the main difference between their work and ours is that we use FNN instead of CNN. Therefore, I've modified the above sentence as "we use FNN on 1D..."

I think there are two main differences:

1. They use CNN instead of FNN (as you say, and it is now clear in the previous bullet point).
2. They tackle the inverse geoid problem directly using ML, while we focus on the forward geoid problem as a first step towards solving the inverse problem using the adjoint method. (perhaps this also justifies why they used CNNs).

[amartinhuertas]

I think there are two main differences:

I made additional changes to make this clearer. Note that you can use, e.g., \vspace{-2ex}, to reduce the space below the figures (which i believe is too much).

[amartinhuertas]

Another concern: the filez size of the poster is now 5MBytes. Is there any file size limit in the submission from Wattle?

[amartinhuertas]

More comments:

Figure 9 and 10.

Title of the figures in the middle. "coefficient"->"coefficients".

Title of the figures on the right. "geoid field" -> "geoid fields".

[GiteonCaulfied]

Hi @amartinhuertas,

More comments:

Figure 9 and 10.

Title of the figures in the middle. "coefficient"->"coefficients".

Title of the figures on the right. "geoid field" -> "geoid fields".

These has been fixed.

Another concern: the filez size of the poster is now 5MBytes. Is there any file size limit in the submission from Wattle?

I've uploaded a copy of the poster on Wattle for testing and no problems occurs, so it should be fine

[amartinhuertas]

why "in essence’s" and not just "in essence" ?

[GiteonCaulfied]

Hi @amartinhuertas,

why "in essence’s" and not just "in essence" ?

I just checked the git log and it seems that this part is actually modified by you in commit 6d6e81a one hour ago (I previously use "in geosciences" and you changed that to "in essence’s"). I've fixed it now (change to "in essence") so it should be fine.

[amartinhuertas]

I just checked the git log and it seems that this part is actually modified by you in commit https://github.com/GiteonCaulfied/COMP4560_stokes_ml_project/commit/6d6e81a0eec2b6567d8540466958d6b42fe9da26 one hour ago (I previously use "in geosciences" and you changed that to "in essence’s"). I've fixed it now (change to "in essence") so it should be fine.

Ok. It should be "geosciences", sorry. It was myself wrongly using the auto-spell checker. Fixed.

[amartinhuertas]

I just pushed last second changes ...

[amartinhuertas]

More comments:

Figure 11. I would replace "above" by the particular data set.

[GiteonCaulfied]

Hi @amartinhuertas

More comments:

Figure 11. I would replace "above" by the particular data set.

This is now fixed.

[GiteonCaulfied]

Hi @rhyshawkins @amartinhuertas @sghelichkhani ,

For the records, the latest version of the poster is now submitted to Wattle.

[rhyshawkins]

Well done, looks good.

[amartinhuertas]

Well done! Now we should focus on the report, which btw, is much more than only 10% of your final mark.

[GiteonCaulfied]

Hi @amartinhuertas @rhyshawkins,

I've uploaded an updated version of the report. All the missing parts in the previous report have been filled.

[amartinhuertas]

As a general comment, I suggest that you look at the output provided by the latex compiler on screen from time to time.

Looking there I could see the warning:

LaTeX Warning: There were multiply-defined labels.

which needs to be fixed.

[GiteonCaulfied]

Hi @amartinhuertas,

As a general comment, I suggest that you look at the output provided by the latex compiler on screen from time to time.

Looking there I could see the warning:

LaTeX Warning: There were multiply-defined labels.

This has been fixed.

Also, I've now made this repository public and refer to it in my report in the Acknowledgement chapter, as we discussed in our previous meeting.