rdf-krr-dos's Introduction

RDF ---> Density of States (DOS) at Fermi level.

A machine learning project to predict DOS at fermi level of sp and spd systems using RDF as descriptor. This project is inspired by: https://journals.aps.org/prb/abstract/10.1103/PhysRevB.89.205118

Descriptor: Radial distribution function (RDF)

Prediction: DOS at Fermi level

1. Create RDF code

A RDF code (RDF.py) is created and tested with XCl (X = Li, Na, K) compounds. Check out the result here: https://github.com/yanxon/RDF/blob/master/XCl_test.ipynb

2. Machine learning procedure

https://github.com/yanxon/RDF-DOS-KRR/blob/master/RDF_DOS_KRR.py (RDF ---> DOS, Aflow)

This script extracts sp metals from AFLOW database. The sp metals are split into train and test datasets. Finally, KRR algorithm is used to train the train dataset and tested with the test dataset.

3. Result

sp metals:

cap: MAE = 0.0022 states/eV/A^3; r^2 = 0.6540

Comparison to PRB (spd system):

cap: MAE = 0.01258 states/eV/A^3; r^2 = 0.6748

rdf-krr-dos's People

Contributors

Stargazers

rdf-krr-dos's Issues

directory to store data_set

@yanxon
In order to let people can quickly reproduce your results.

I suggest you create a directory called source_data, in which you download and dump these data in json format.
For instance,
sp_metal_aflow_844.json
For each structure, you save the following information,

structure - lattice, coord, atom array
property - dos, band-gap, formation energy, .etc.

formation energy

@yanxon

Please try to predict the formation energy with the Jarvis data set.

I suggest you reorganize the files in the following structure

root_dir
├── Descriptors (RDF.py, ADF.py, ... etc)
├── Datasets (some json files)
├── Results (store results, such as png files)
├── RDF_DOS.py
├── RDF_Formation_energy.py

This can greatly improve the readability.

Recommend Projects