Machine Learing¶
In [3]:
import pandas as pd
import numpy as np
from matplotlib import pyplot as plt
import seaborn as sns
from pymatgen import Structure, Lattice
from pymatgen.core import periodic_table
from pymatgen.analysis import ewald
%run read_struct_get_ewald.ipynb
%run read_struct_get_bond.ipynb
%run read_struct_get_coordination.ipynb
%run read_struct_get_coulomb.ipynb
from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor
from sklearn.model_selection import ShuffleSplit, cross_val_score, train_test_split
from sklearn.preprocessing import Imputer
In [3]:
train = pd.read_csv('train5.csv')
test = pd.read_csv('test5.csv')
In [342]:
train.head()
Out[342]:
In [343]:
train.columns
Out[343]:
Random Forest/Gradient Boosting Regression¶
In [345]:
features_bandgap = ['mass_al','mass_ga','mass_in','mass_tot','vol','spacegroup', 'percent_atom_al', 'percent_atom_in','density','z_al','z_in','z_tot','inv_r_alo_sum','calo','inv_r_ino_sum','cino','inv2_r_sum','r_min','rmean','nalo_mean','nino_mean','n_mean',
'coul','calal','cinin','cgaal','cgain','calin']
features_formation = ['mass_al','mass_ga','mass_in','vol','spacegroup', 'percent_atom_ga','z_ga','inv_r_gao_sum','cgao','r_min','ngao_mean','n_mean',
'cgaga','calal','cinin','cgaal','cgain','bandgap_energy_ev']
print('Bandgap features = ', features_bandgap)
print('Formation features = ', features_formation)
Only features that have high correlation with the bandgap/formation energy are chosen.
In [11]:
def ml_regression(train,test,x_feature,y_feature,model,feature_importance=True):
x_train = train[x_feature].values
y_train = train[y_feature].values
x_test = test[x_feature].values
print('Defining RFR model...\n')
model.fit(x_train,y_train)
score = model.score(x_train,y_train)
print('Training Score = ', score)
# CV score
cv = ShuffleSplit(len(y_train), n_iter=10, test_size=0.1, random_state=0)
cvscore = cross_val_score(model, x_train, y_train, cv=cv, scoring='r2')
print('CV Score = ', cvscore)
print('Mean = ', np.mean(cvscore))
predict = model.predict(x_test)
if feature_importance:
importances = pd.DataFrame({'feature':x_feature,'importance':np.round(model.feature_importances_,5)})
importances = importances.sort_values('importance',ascending=False).set_index('feature')
print(importances)
importances.plot.bar()
plt.show()
return model, predict
In [328]:
def submission(predict_bandgap,predict_formation):
submission = pd.read_csv('sample_submission.csv')
submission['bandgap_energy_ev'] = predict_bandgap
submission['formation_energy_ev_natom'] = predict_formation
submission.to_csv('submission.csv',index=False)
In [12]:
def split_train(train,x_feature,y_feature,model):
train_2, test_2 = train_test_split(train ,train_size = 0.8)
x_train_2 = train_2[x_feature].values
y_train_2 = train_2[y_feature].values
x_test_2 = test_2[x_feature].values
y_test_2 = test_2[y_feature].values
model.fit(x_train_2,y_train_2)
predict_2 = model.predict(x_test_2)
plt.scatter(y_test_2,predict_2)
plt.plot([0,train_2[y_feature].values.max()],[0,train_2[y_feature].values.max()],'--')
plt.xlabel('Actual')
plt.ylabel('Predict')
return plt.show()
In [348]:
print('Bandap Energy:')
model_bandgap, predict_bandgap = ml_regression(train,test,features_bandgap,'bandgap_energy_ev',RandomForestRegressor(n_estimators=1000))
#model_bandgap, predict_bandgap = ml_regression(train,test,features_bandgap,'bandgap_energy_ev',GradientBoostingRegressor(n_estimators=100))
test['bandgap_energy_ev'] = pd.DataFrame(predict_bandgap)
print('Formation Energy')
#model_formation, predict_formation = ml_regression(train,test,features_formation,'formation_energy_ev_natom',RandomForestRegressor(n_estimators=100))
model_formation, predict_formation = ml_regression(train,test,features_formation,'formation_energy_ev_natom',GradientBoostingRegressor(n_estimators=500))
Prediction scores (R2) on Cross-Validation are 95.0% and 89.4% on bandgap and formation energy. Notice Gradient Boosting Regression is used on formation energy prediction, because the Random Forest Regression gives high bias on traning score. Predicted bandgap is used as a feature to predict the formation energy due to their high correlation.
In [331]:
submission(predict_bandgap,predict_formation)
In [353]:
split_train(train,features_bandgap,'bandgap_energy_ev',RandomForestRegressor(n_estimators=1000))
split_train(train,features_formation,'formation_energy_ev_natom',GradientBoostingRegressor(n_estimators=500))
In [ ]: