In a typical NN training, RMSE(E) is of the order of 1-2 meV per atom and RMSE(F) of the order of about 0.1 eV/Å. [2020 - Machine Learning Meets Quantum Physics pp 253-275, High-Dimensional Neural Network Potentials for Atomistic Simulations]

SchNet

• Atom-wise fully connected layers: $\textbf{x}_i^{l+1} = W^{(l)}\textbf{x}_i^l + \textbf{b}^l$
• The softplus activation is chosen due to its high-order differentiability which is crucial for the prediction of atomic forces and/or vibrational frequencies.