What I did

• Monday: Got FEFF to work. Read paper: “Probing Atomic Distributions in Mono- and Bimetallic Nanoparticles by Supervised Machine Learning” Nano Lett. 2019.
• Tuesday: Ran the simulations for FEFF on range -.45 to +.45 \(\Delta\rho\) shift and looked at the results.
• Wednesday: Thought about problems 1 and 2.
• Thursday: Thought about problem 2. Came to conclusion.
• Friday: Ran Feff simulations successfully
• Saturday: Compared the feff spectra with the new parameters

What I learned

• Monday:

  • Windows .exe FEFF files work fine and you can run bash scripts in powershell just by typing bash to enter linux like mode and running calling ./file.sh to run a script like you normally would in linux. Note, you can call bash scripts to execute .exe files in windows, but if you put the .exe files in your WSL directory and then try and call them, it won’t work.

  • In the paper, there were a bunch of RDF and PRDF (Partial Radial Distribution Functions). The standard deviation of the PRDF is what I’m considering as sigma in my simulations. In the paper, the STD of g(x) for the Au atoms in the RDF tended to be around .1 Angstroms.

• Tuesday: There is a change in the weighted (disordered) spectrum from the original, unweighted and unshifted spectrum. Two problems: 1. FEFF won’t generate disorder from a shift as large as \(\Delta\rho=+0.45\). I’m trying to come up with a solution where I can replicated the PRD without having to do RMC MD simulations. Perhaps I can just weight with a left tailed spectrum … well left tailed in the sense that I would just be neglecting the right tail of a normal distribution. And 2. I’m a little concerned that the difference between gaussian weighted spectra might be too small for the sensitivity of a NN. Maybe not. I’m not as concerned with that problem as problem 1.
• Wednesday: I thought about having way more spectra within the range of \(\pm0.25\). I think maybe I could have more points, but 10X would be overkill. I don’t think you can replicate that PRDF with a different gaussian scaling. But maybe you can have all the possible sigmas weight the spectrum by only taking into account the inner 90% or something of the spectra…whatever the fattest gaussian will allow.
• By changing to a fixed height gaussian for ALL spectra, weighting all with only 90% of the area of the gaussian should probably still work. Would be nice to have
• Friday: Shifting CS value from .2 -> .05 and the SCF Radius from 3.1 to 3.5.
• Saturday: uh oh they look pretty different. Need to compare with previous data to see if the new values are valid.

What I will do next

• Investigate if I can change the parameters and get simulations that compare better to actual data.