xnn.ffnn#

ffnn: learnable classical force fields.

Classical force-field functional forms whose parameters (and, optionally, small embedded neural networks) are trainable by gradient descent – the force field is the model. The first member is ReaxFF, the bond-order reactive force field of van Duin et al. (2001), together with its machine-learned variant ReaxFF-nn (Guo et al. 2020, Xue et al. 2021). The second is OPLS, the fixed-topology all-atom/united-atom force field of Jorgensen et al. (1996), covering its variants (OPLS-AA, OPLS-UA, L-OPLS) through interchangeable parameter libraries.

Parameter libraries for the whole family are read from the MolSSI/SEAMM .frc force-field format (xnn.ffnn.common.frc), which also carries the SMARTS templates that type a structure (xnn.ffnn.common.typing); the OPLS-AA distribution and a dozen published ReaxFF fields ship with xnn (list_forcefields()).

Importing this package registers the family’s models with the shared model registry (xnn.common.models).

Modules

common

Shared infrastructure of the ffnn (classical force field) family.

models