TODO: update this readme Critical role descriptor design doc for generating the thousands of columns:
'''for each element in elem site order: find the index in elem by formula order use this index to scrape each atom_rad till num electrons, then add them to a dict in that site
for me: A1 (aka P1) will be atom radius, atom volume,
A2 will be dipole polarizability, electron affinity
A3 will be covalent radius, num electrons
in order to read as floats and ints, I will store the method of finding the value as the label of the column
starting out will stick to A-A, A-X, B-X interactions
here are the calculations to see how many columns will be generated NOTE: IDs from the paper are ABCDEF, since this will be confusing with A and B sites, I will call A and B ids P and Q now: P1: 10 Atom radius and Atom volume: number of features is 10 (2 * 5 ) (atom rad * 5 of a1a2b1b2x, atom vol * 5) P2: 10 Dipole Polarizability, Electron affinity P3: 10 covalent radius, num electrons
the paper is 4,4,6, so I have more features to start with. to prevent this from beign too large an issue I will start by only combining the following:
NOTE: the combinations in the paper are all combinations of two different sites (no A A or B B), with two different symbols to connect them (+ or -). there are overlaps that are not included (A+B = B+A). this is how they are generated, it took an embarrasing amount of time to verify that
here are the interactions I will consider:
f(A1, A2), f(A2,A1) f(A1,X) f(A2,X) f(X,A1) f(X,A2) f(B1,X) f(B2,X) f(X,B1) f(X,B2) there are + and - for all of them, which makes the total number for Q1 20, -5 for a1+a2=a2+a1 = 15 This is slightly higher than the 12 from the paper, but manageable
as an example, one of the Q1 descriptors is something like this: ar = atom radius, av = atom volume
|ar(A1)+ar(A2)|
Q1: absolute differences and sums of P1 :
this must be very modular, since I am creating order 10,000 columns.
plan of action: make list of the 10 interactions that I can loop through (f(A1, A2), f(A2,A1) f(A1,X) f(A2,X) f(X,A1) f(X,A2) f(B1,X) f(B2,X) f(X,B1) f(X,B2))
create a function that takes the two site names as arguments, finds their 2 interactions, and appends them to the respective columns (must compute column name)
calculate(A1, A2, abs_sumdiff, dataframe)
data stored like so: column name = "P2+", "P2-"
inside column store dict of key value, where key is "ipA1" or "ipA1-ipA2"
'''