LASP2 Documentation
LASP2 is an interface for performing atomic simulations using a High Dimensional Neural Network Potential (HDNNP) with on-the-fly training. LASP2 is intended to be a bridge between the molecular dynamics software LAMMPS, the HDNNP software n2p2, and the DFT software VASP. LASP2 runs a LAMMPS simulation using a HDNNP potential generated by n2p2. When training is needed, the simulation is stopped and a DFT calculation is performed using VASP, this is then used to extend the database and train the potential with n2p2.
Main concept
An HDNNP is used to predict the forces experienced by the atoms in the system, based on a reference method (DFT in this case). The forces predicted will depend on the values of the weights and biases of the neural network, which are in general optimized during the training procedure. However, when the HDNNP encounters a structure with unknown characteristics the result will strongly depend on the initial values of the weights and biases before the optimization.
The main assumption of LASP2 is that an HDNNP will produce random results when predicting forces for an atomic structure that was not part of the training database. We take advantage of this by training HDNNP's with different random seeds and using the dispersion as an indication of similarity to the training database. When the dispersion increases over a defined thershold it can be said that the result is at risk of not having physical meaning, therefore a DFT calculation is performed for that structure and included in the training database.
Using this method, the goal of LASP2 is to perform simulations at speeds similar to atomic-based methods while occasionally stopping to perform expensive DFT calculations when necessary in order to approximate the accuracy of electronic-based methods.
Reference
Created: August 9, 2022