Add class method for Box class which create a box from Lx,Ly,Lz and angles #1234
Conversation
for more information, see https://pre-commit.ci
…tzerlab/freud into feat_from_to_lattice_vectors
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
…tzerlab/freud into feat_from_to_lattice_vectors
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
|
I personally am a proponent of using math.sqrt and math.acos, as for single values it is significantly faster than the numpy alternatives. The difference is on the order of μs, but free performance is free imo. |
It seems numpy sqrt is used throughout freud from what I can tell. If we were to start using math functions we should change all occurrences. I agree it makes no sense to leave performance on the table, but realistically the differences will be so small they are inconsequential. Perhaps this would be a good first issue to tackle for someone new? |
freud/box.pyx
Outdated
| v2 = lattice_matrix[2] | ||
| Lx = np.sqrt(np.dot(v0, v0)) | ||
| a2x = np.dot(v0, v1) / Lx | ||
| Ly = np.sqrt(np.dot(v1, v1) - a2x * a2x) |
There was a problem hiding this comment.
| Ly = np.sqrt(np.dot(v1, v1) - a2x * a2x) | |
| Ly = np.sqrt(np.dot(v1, v1) - a2x**2) |
I think a square here better conveys the intent, as below we have a very similar function with two different variables (a2x * a3x)
Noted, and sounds good! I will raise an issue. Edit: See #1236 |
Co-authored-by: Tommy Waltmann <[email protected]>
tests/test_box_Box.py
Outdated
| 1, | ||
| 2, | ||
| 3, | ||
| np.pi / 3, | ||
| np.pi / 2.25, | ||
| np.pi / 2.35, |
There was a problem hiding this comment.
Each of the six numbers should be the output of a different call to np.random.rand. That way, each time the test is run it gets run with a different random value.
There was a problem hiding this comment.
I am currently working on adding this.
for more information, see https://pre-commit.ci
freud/box.pyx
Outdated
| a2 = np.array([L2 * np.cos(gamma), L2 * np.sin(gamma), 0]) | ||
| a3x = np.cos(beta) | ||
| a3y = (np.cos(alpha) - np.cos(beta) * np.cos(gamma)) / np.sin(gamma) | ||
| a3z = np.sqrt(1 - a3x**2 - a3y**2) |
There was a problem hiding this comment.
This line appears to take the square root of a negative number at times, resulting in NaNs.
There was a problem hiding this comment.
some combinations of angles are not possible-there is likely a constraint on their sum or something like that. Is this taken into account?
There was a problem hiding this comment.
Not in how I wrote the test case, which would explain the failures. The code should throw an error if the user gives inputs which violate the constraints.
|
All looks good to me, just one comment. If we want to just use freud for the EBT box, it would be good to have a convenience function converting to a unit vector + lengths description. Do we feel that this would fit into freud, or should this just be separately implemented in ebonds? |
I would start by adding the functionality into Right now, the best way to get lengths and unit vectors would be to get the box vectors (they are the columns of |
|
|
It's fine to leave atol as 0.0. According to the documentation for |
A non-zero |
|
I have tested this method against box lengths and angles from 1,104 real CIF files and all cases pass. I can export the box params as a numpy array and save it as test data, or we can adjust the current test to be more realistic. |
The code should be written so invalid sets of angle inputs raise an error to the user instead of creating invalid boxes. This PR needs an update to the code, not just the tests, before it's ready to be merged. Ideally, we would know the mathematical constraints on the box angles so we could check that the inputs are valid, and we should make an effort to figure out what those are. If we cannot get those relations, then we could move forward with an updated set of tests and a "garbage in, garbage out" note in the documentation, but again, I only think we should take the second option if we have already given the first option some effort and failed. |
…for angles and raise valueerror when violated
|
I think I figured it out. The restriction is the following: The geometric interpretation of this is the violation of the cosine law. |
Implements classmethods that create a box class from
lattice vector matrix andL1, L2, L3 and angles, as well as methods that give L1, L2, L3 and angles from box class.Description
Crystals are usually defined via their lattice vectors and wykoff sites or lengths of lattice vectors and angles; while freud and hoomd use cell parameter matrix. In this PR I implement class method for creation of box
from lattice vectors orfrom lattice vector lengths and angles. I've also implemented a method that gives lengths of box vectors and angles from box object.Motivation and Context
Lattice vectors or lattice vector lengths and angles are used to define real crystals and having a method that works with this makes creating crystals easier.
How Has This Been Tested?
Tests were added for new methods for box testing.
Types of changes
Checklist: