General questions

[Wenhao-Shao]

Hello Ziyang,

I hope this message finds you well. This is Wenhao Shao. I recently started to implement the ReflectoNet from you group. It worked well so far and the default setting printed

400
0 0
0.43365893626201923 (train_distant_n)
0.20817030273284456 (train_distant_k)
crosscoef: test Accn=0.7225
crosscoef: test Acck=0.9479

and an image, which I don't know the exact meaning.

May I kindly ask a few clarifications?

1. Is the code and most up-to-date?
2. What is modeltester.py doing exactly and what is the output image?
3. I see that it's currently taking in "./raw_data/blindtestinput.csv" and "./raw_data/blindtestoutput.csv", and outputing "blindpredicted.csv". How to read these csv files? (e.g. what is the physical meaning of each row and column? what is blindtestinput and blindtestoutput exactly?)
4. If I wanted to use my own reflection spectra, how should I adjust the input for the prediction of wavelength-dependent n and k?

Thank you and I look forward to hearing from you!

Best,
Wenhao

[tigerwang3133]

Dear Wenhao,

Thank you for your interest in the code!

Here are the answers

1. Yes, it is the current code. In the generated figure: green is ground truth refractive index; red is predicted refractive index by model. The first 400 values are n in 400nm-800nm. The next 400 values are k in 400nm-800nm.
2. The modeltester.py is the demo for reading reflection and predicting the corresponding refractive index with the model. Then, the code compares the prediction to the ground truth refractive index.
3. /raw_data/blindtestinput is the input reflections. Each row is the 5 sets of reflections, (target material with substrate, substrate only)*5. Each length of the reflection is 400, representing wavelength from 400nm-800nm.
   /raw_data/blindtestoutput is the ground truth of the refractive index. Each row is one set of refractive index (n, k). Each length of the n or k is 400, representing wavelength from 400nm-800nm.
4. If you want to use your reflection data, you can size your reflection into 400nm-800nm with 1nm interval, length 400. Then rearrange and flatten 5 sets of reflections into (target material with substrate, substrate only)*5. Then check the predicted values by the model.

Please let me know if you have any questions. You can also email me at zw76@rice.edu

Yours sincerely,
Ziyang Wang

[Wenhao-Shao]

Dear Ziyang,

Thank you so much for your prompt response and for keeping the code well maintained! Your clarifications made everything much clearer, and I just have a few quick follow-up questions. I’ll post them here in case others would like to take a look as well. Thanks for sharing your email—I appreciate it. I'll be sure to reach out if I have more specific questions in the future.

1. What is the nature of each column in blindtestoutput.csv or blindpredicted.csv? Are they from 400 distinct materials/materials systems? I would assume that blindtestoutput.csv is the ground truth and you are comparing the first column with that in blindpredicted.csv to generate the figure, but please let me know if I'm wrong.

2. I understand from the paper that a set of two reflection spectra of (target material with substrate, substrate only) is needed for prediction. Given that you recommended 5 sets for each targeted material, is this just for duplication, such as measure 5 different flakes of 2D material?

3. blindtestinput.csv contains 4000 columns. Are they like "400 materials systems" x 2 spectra in the set (with material + bare substrate) x 5 sets?

4. Just for curiosity, blindtestmateriallist.csv seems to have 2000 materials systems. Is the sheet just there as a note and/or is it used in the model "0.1model.pt"?

Again, thank you for your kind assistance!

Best,
Wenhao

[tigerwang3133]

Hello Wenhao,

1. Each row is one set of distinct materials/materials systems. I am comparing the first row in the blindpredicted.csv to blindtestoutput.csv
2. Yes, I recommend 5 sets for the same target 2D material with different structures of substrates (different thicknesses or different substrate material). This is to make sure the variance of the reflections is large enough to provide information on the target material. You can try to change the order of the 5 sets or change the number of inputs (putting duplicate material sets)
3. Yes your are right, blindtestinput.csv contains "400 materials systems" x 2 spectra in the set (with material + bare substrate) x 5 sets. The input of the model will be the 5 sets of material systems.
4. These are the notes for generating this blind test from the material in the C2DB database. They are not used in the training of the model.

Yours sincerely,
Ziyang Wang

[Wenhao-Shao]

I see! so each row is one set of data instead of each column. I thought it was a typo before. Everything is clear now. Thank you very much!

Wenhao

[Wenhao-Shao]

Hi Ziyang,

May I follow up and ask how to use my own measures reflection spectra for prediction? After flatting it into a row of 4000 datapoints (400x2x5) as you suggested, would you replace the original blindtestinput.csv now with the new csv file containing the new data? This doesn't seem to work since originally, blindtestout.csv was used as the ground truth (labels).

Thank you for you assistance in advance.

Wenhao

[tigerwang3133]

Hi Wenhao,

You can replace the original blindtestinput.csv with your new csv. Then, you can replace blindtestout.csv with a random row of 800 data points since you do not have the ground truth to compare.

Ziyang Wang

[Wenhao-Shao]

Hello Ziyang,

I hope this message finds you well. I tested the model using my samples, which are exfoliated 2D perovskite PEA2PbI4 on 500nm SiO2/Si. 5 curves were measured with different flakes and all flattened to be paired with the same substrate curve.

I found two problems with the prediction:

1. the predicted n and k dispersions are very sensitive to the scale of the input data in the y axis (intensity). The raw data was in the scale of 100-10k, and I had to normalize them to 0.01-1 in order for the predicted n and k to make sense. A scale that was too low or too high both did not work. You can find this in my note attached.
2. The predicted peak near 520nm for n and 495nm for k kind of make sense for this material, but the broad band predicted near 670nm, or from 1.5-2.0eV does not. This could come from artifacts from interference as the sample flakes may not be thin. I was wondering if you have seen this before. Any suggestions are appreciated!
3. The absolute values of n or k (even for <550nm or >2.0eV) did not make sense as they seem to be too high. For your reference, here are some reference having the n and k dispersion of this material either measured with ellipsometry (https://opg.optica.org/oe/fulltext.cfm?uri=oe-32-14-25327&id=552913) or simulated (https://pubs.acs.org/doi/10.1021/acsomega.4c01827). I appreciated it if you have any insights on this.

Here is a note I wrote summarizing the results and raw data.
Testing of ReflectoNet 25.4.docx
Here is the raw data, unscaled.
data.xlsx

Any suggestions or insights are much appreciated! Thank you for your time.

Best,
Wenhao

[tigerwang3133]

Hello Wenhao,

1. For the input, you need to use the relative reflection that scaled 0-1. For example, you can measure the reflection of a silver mirror to get intensity R0, then measure the sample, get intensity R1, and then measure the substrate to get intensity R2. Then, the input of the model will be [R1/R0, R2/R0], which is 400+400=800 numbers. You will have 5 of these. Then, you can get input within a scale of 0-1.
2. I see your light spot is not full circle. That could make the measurement less accurate.

[Wenhao-Shao]

Thank you, Ziyang. I'll test again with the relative reflection.

Related to the experiments, what's your opinion on whether the light spot should cover the entire flake, or whether the flake should be large enough to cover the light spot?

Best,
Wenhao

[tigerwang3133]

I believe the flake should be large enough to cover the light spot. This will make sure that your measured reflection is all from the flake.

[Wenhao-Shao]

Got it. Thank you!