Some gradient

[Rafael-G-C]

@bast here's the code that returns the error

TypeError: Gradient only defined for scalar-output functions. Output was [array(0.), array(0.), array(0.), array(0.), array(0.), array(0.), array(0.), array(0.), array(0.), array(0.)].

just run the test as python3 -m pytest tests/ in the pyDFTD3 folder on line 93 in test_CO2H2_2.py "2" is the hessian d3_dervs = D3_derivatives(charges,functional,damping,2,coordinates)

[bast]

And what we want to find out is: why is the output a list instead of a float?

[Rafael-G-C]

Yes, and now that I look at it why it a list zeros

[bast]

I think the code fails because the first element in variables here is an array of charges (see here). I think this is some leftover?

[Rafael-G-C]

Okey, might be let's see what's happens when I fix it

[bast]

Also in the test I would not start by cycling through all derivatives but start only with very few elements. It takes time to cycle through all of them and right now we will be very happy if already 1 or 2 work which takes a fraction.

[Rafael-G-C]

I see so stopping the derivative order loop earlier

[Rafael-G-C]

Alright the code can now: get the gradient, hessian, 3rds and 4rds of the sum of the r6, r8 and repulse interactions in kcal this line derivative_orders = list(derivative_orders)[0:1] is used to cut off the list of derivative orders so we have control of how many we want

[bast]

This sounds fantastic!

[robertodr]

Really great Rafael, apologies for tardiness, I didn't get email notifications for these changes... I'll try to fix the conflicts with the base branch and get a test going.

[robertodr]

OK, so:

• I have rebased on top of master and fixed the conflicts. @Rafael-G-C, before you do any edits locally: git pull --rebase origin some_gradient
• The main function (d3) accepts 3 arguments: the configuration, the list of charges, the list of coordinates. The configuration is a dataclass.
• I have re-enabled the tests for the energy. The reference numbers come from the implementation in LSDALTON.
• I have added tests for the gradient. Again, the reference numbers come from the implementation in LSDALTON, which is analytical, not numerical.
• I have updated the requirements.txt, since we officially depend on JAX now. I dropped Windows in the GitHub action.

[robertodr]

A word of warning: the gradient test is pretty slow.

[robertodr]

The tests should hopefully run slightly faster now (I've removed one of the parametrization levels for the gradient test)

I have also reorganized a bit the code to make it easier (hopefully!) for @ringholm to run the numdiff tests:

1. To run a n order derivative calculation:

   python -m dftd3.dftd3 --damp zero --order 1 examples/formic_acid_dimer.com

2. This outputs the energy to screen and writes a JSON file with the full input and full output (including the gradient). In Python, you can load the gradient with:

   import json
   
   import numpy as np
   
   with open("formic_acid_dimer.json", "r") as f:
          data = json.load(f)
   
   gradient = np.array(data["formic_acid_dimer"]["output"]["1-st order geometric derivative"])

[robertodr]

Merging so that @ringholm can try the numerical differentiation stuff before the next meeting.