Fig. 1: Neural network architectures explored in this work.
From: Teaching a neural network to attach and detach electrons from molecules
Models from literature: a ANI13, b AIMNet8. Here each model is separately trained for neutral species, cations, and ions. Models introduced in this work: c AIMNet-MT: a multitask model jointly trained on all data which concurrently predicts energies and charges for neutral species as well as cations and ions; and d AIMNet-NSE, a Neural Charge Equilibration model which is capable to re-distribute spin-polarized atomic charges according to a given molecular spin charges and predicts energy for the specified (arbitrary) spin state of the molecule. The yellow blocks show input data (coordinates R, atomic numbers Z, and total molecular spin charge Q) and output quantities (energies E and spin-polarized charges q). The green blocks denote trainable modules, and the blue blocks are fixed encodings.
