Universal Artificial Neural Network Model For Accurate Band-gap Prediction Of Main-group Alloys

The band gap is a significant electronic property to be concerned with the fabrication of semiconductor devices.By controlling their composition and concentration,alloys can get rich of band gaps satisfy the needs of different use.But acquiring band gaps of alloys quickly and accurately are difficult in experiment or by first-principle calculations despite the developed computational power of computers and experimental facilities.Precise prediction of material property is received increasing attention with the assist of machine learning methods in the Materials Informatics and always need more accurate data and appropriate feature engineering in the process.We construct zinc-blende(ZB)and wurtzite(WZ)database of binary and ternary alloys based on main group element through first-principle calculation filling in the vacancy in the existing database and design artificial neural network(ANN)model that trained with a data set combining variable-fidelity first-principle calculation results of alloys band gaps enable prediction of the band gap at the high fidelity level.After training with a set of 1559 ZB alloys as an example data set,the ZB multi-fidelity model effectively in predict high fidelity band gaps of ternary alloys with the coefficient of determination(R~2)scores over 99%and the mean absolute error(MAE)below 0.09 e V,then we get ternary alloys high fidelity band gaps with continuous change of mixing ratio.Furthermore,we extend our ZB multi-fidelity model to predict quaternary alloys shows the model generalization with reasonable result.Finally,3335WZ alloys are trained to get WZ low fidelity model and then applied to the multi-fidelity ZB model bypass the time-consuming high fidelity band gaps first-principle calculation with capable of predict WZ high fidelity band gaps with different space groups directly.We introduce two different universal features construction methods and indicate simple one-hot-encoding representations of elements are comparable to traditional elemental properties representations be handled with careful manual selections in the process.And importing the low fidelity band gaps to ANN model improve the prediction precise remarkablely.Our model is naturally fitting the relationship between the band gap and the continuous elemental fraction change in the ternary alloys and can extend to more multiple alloys with different configurations achieving high fidelity band gaps prediction with suitable results.The presented work flow framework is not only applied to the band gap prediction,but also convenient applied to the utility of property prediction in various aspects as a universal ANN model.