Fig. 1: Graphical representation of the ensemble-scope descriptor.

The present scheme is composed of three independent machine learning models for the oxygen-ion conductivity (σ) as the target property. The first model employs a knowledge-based handcrafted descriptor whose dimensionality is optimally reduced by the PLS model. The second descriptor is SOAP, which calculates a metric between compounds with respect to the atomic geometry and electronic negativity. The SOAP is used for the kernel-ridge regression model. The third descriptor is R3DVS. This descriptor encodes the BVS voxel data, which approximate the oxygen-ion transportation path into a fixed-length array that is suitable for the 3D-CNN model.