Fig. 6: The Deep Set architecture. | npj Computational Materials

Fig. 6: The Deep Set architecture.

From: From individual elements to macroscopic materials: in search of new superconductors via machine learning

Fig. 6: The Deep Set architecture.

A schematic layout of the Deep Set architecture here employed. The input of the Deep Set is the set of elements that compose the analyzed compound. Each element is associated to 22 chemical/physical instances as described in the main body of the paper. The stoichiometric contribution of a given element defines an additional entry of the supplied vector. The input is processed by the non linear function ϕ, represented as a feed forward neural network. The result of the processing are gathered together and further transformed by the application of a non linear filter ρ. This is also defined in terms of a neural network. The produced output is confronted to the expected target, during training stages, and the parameters of the networks ϕ, ρ tuned so as to minimise the assigned loss function.

Back to article page