Fig. 11: Network intermediate output (activation) inspection example.

The authors predict material property from elemental composition inputs with a fully connected neural network. They then examine network activations by compressing the activations of different materials with PCA and plotting the first two PCs. Results show that intermediate network layers can learn essential chemical information. Early network layers tend to learn the presence of elements and later network layers tend to learn the interaction between elements (e.g., charge balance). Figure reprinted from ref. ⁸⁹ under the CC-BY 4.0 license¹⁵⁶.