Machine-Learned Chemical Pressure Analysis
This repository is divided into two directories. In the "ML-CP training" directory, you will find the code and data used to produce the two models discussed in our ML-CP preprint. The models are exported as .pickle files. With these pickle files, you can use the code and data contained in "trained ML-CP model" directory to make ML-CP schemes.
All code is written in python, and may require python version 3.8 and SciKit Learn version 1.3.2 to run. At the very least, the training and utility code will need the same version of python and SciKit Learn.
To run the machine_learning.py code to produce the classifier and regressor models, download the files contact_data_by_cp.xlsx, element_data.xlsx, and machine_learning.py from the "ML-CP training" GitHub directory. With these three files in the same local directory, run the command
python3.8 machine_learning.py
This operation should take 1-2 minutes, and produces two new files, rfc.pickle and rfr.pickle. Now to produce ML-CP data, copy the resulting .pickle files to a new local directory containing the element_data.txt and mlcp.py files from the "trained ML-CP model" GitHub directory. You'll also need to include in this local directory a Crystallographic Information File (CIF file) for the structure you're interested in producing ML-CP data for. With the two .pickle files, the element_data.txt file, the mlcp.py file, and the CIF file in the same directory, run the command
python3.8 mlcp.py [CIF filename]
Optionally, you can also provide a scale factor and template. Scale factors can be used to produce a ML-CP scheme with a total CP closer to 0 by artificially scalling the a, b, and c cell lengths. subcell templates can be used to produce ML-CP data for subsets of the unit cell. This functionallity allows the user to produce ML-CP data for regions of a unit cell with lower requirements for memory allocation and time. The mlcp.py program currently only accepts XYZ files, which can be produced using VESTA or other structure viewing programs. To run the mlcp.py program with a scale factor, run the command
python3.8 mlcp.py [CIF filename] [scale factor]
To run the mlcp.py program with a scale factor and template or just a template, run the commands
python3.8 mlcp.py [CIF filename] [scale factor] [template filename]
or
python3.8 mlcp.py [CIF filename] 1 [template filename]
In the latter case a scale factor is still required, so 1 is used as a placeholder without modifiying the scale. Please note, if a template is used, it will not be possible to calculate the net CP of the cell, so that information will not be outputted. Regardless of if the scale factor or template options are used, the mlcp.py program will create five new files, following the format [CIF filename]_MLCP-[file type]. The five file types are:
- -cell
- -geo
- -coeff
- -bptable
- -data
The first two describe the unit cell and atomic positions of each atom, both in cartesian coordinates. The -coeff file contains coeffecients for the l=0 - l=4 spherical harmonics. The -bptable file contains data about each interatomic contact in the cell, including the identity and coordinates of each atom in the contact, the contact length, the average pressure (in units of eV/Å3), the number of voxels in the contact, and the volume of the contact in Å3. The -data file contains information about the ML-CP calculation, including the CIF filename, the scale factor used, the cell parameters, a CP quadrupole report, the net CP of each atom in the cell, the net CP of the cell (if a template was not used), and the calculation time.