We use the recently proposed ML method (J. Phys. Chem. A 127 (2023) 7823) that has the form of a single hidden layer NN that uses additive Gaussian process regression to construct optimal neuron activation functions, to machine learn the band gaps and heats of formation of lead-free inorganic halide double perovskites that are promising for solar cell applications. The method combines the high expressive power of a neural network with the robustness of a linear regression. It avoids non-linear optimization that is largely responsible for the CPU cost of an NN and overfitting. We show that better prediction quality can be obtained, in particular for the bandgap in the visible region relevant for most applications, compared to previous results using standard methods. The model does not suffer from overfitting as the number of neurons is increased, in spite of the low data density in a relatively high-dimensional feature space. Our method also helps identify the most important features among about 30 descriptors of chemical composition and structure, and the importance of coupling among features.
Sergei Manzhos