NEP-Charge

[brucefan1983]

Summary

Add the charge degree of freedom into the NEP approach. The resulting method will be called NEP-Charge.

• Usage in nep.in:

has_charge 1 # default is 0 (check your screen output)

• Formalism:

  • Similar to this: Charge-Optimized Electrostatic Interaction Atom-Centered Neural Network Algorithm [1]
  • Learn the global charge of the structure only, not partial charges on the atoms. Ref. [1] does not learn the global charge. Besides this, charge neutrality is enforced.
  • Charge and potential share most of the neural network but have different weights from the hidden layer to the output layer. Using completely separate neural networks is not benificial.
  • Electrostatic interactions are evaluated using the standard Ewald summation. The speed is ok for training but I need to explore more efficient algorithms for MD simulation.
  • Virial and heat current are carefully derived. Still very elegant formulation, in the spirit of Force and heat current formulas for many-body potentials in molecular dynamics simulations with applications to thermal conductivity calculations. These are our new contributions.
  • It was suggested that the real-space and self-energy parts could be removed Latent Ewald summation for machine learning of long-range interactions, but I prefer to keep them, to ensure the integratiy of the charge degree of freedom.

• A typical charge distribution predicted for the PbTe example:

• Force checked against my Matlab (no chance to be wrong) code:

• To solve the issue adding charges to NEP #299
• TODO: documentation + MD part