In this work, we offered a machine-learning based predictive model that output the phonon density-of-states by inputting atomic structures. One feature is the prediction of alloy properties. Conventionally alloy prepdictions can be achieved using virtual crystal approximation (VCA), where for the an alloy AxB1-x, the virtual mass mVCA and potential VVCA are given by
mVCA = x * mA + (1-x) * mB
VVCA = x * VA + (1-x) * VB
In the machine-learning model, using one-hot encoding to represent one pure element with certain mass (i.e. H = [1,0,0,…], He=[0,4,0,0,…]), the feator vector for alloy AxB1-x can be written as
[0,0,…., x * mA, 0,0,…, (1-x) * mB,0..]
So, two VCA equations emerged into one for machine learning, but without generating additional computational cost when computing alloy properties.