Materials Property Prediction
Materials property prediction uses statistical or machine learning models trained on computed or measured data (often high-throughput DFT databases such as the Materials Project) to estimate properties such as band gaps, formation energies or elastic moduli of new materials without running a full calculation. The choice of descriptors (inputs) controls both accuracy and whether the model can be inverted to design materials with target properties.
Kernel regression on composition and space group (2016)
Calfa and Kitchin described each crystal by only its space group number and the count of each element in its formula, and fitted nonparametric kernel regression models with a Racine-Li kernel for categorical data, with one bandwidth per predictor chosen by leave-one-out cross-validation (Calfa 2016). On Materials Project data the models reproduced band gaps of 746 binary metal oxides with an RMSE of about 0.05 eV and Voigt and Reuss bulk and shear moduli of 1173 crystals with RMSEs of about 0.7 to 2.2 GPa; the Fermi level (MAPE about 10%) and the total energy of the chemically diverse elastic set (MAPE about 7%) were the weakest. The descriptors were chosen, by analogy with group contribution methods, so that a prediction maps directly back to a composition and structure. Summary: calfa2016.