Non-conservative force

Non-conservative forces are forces that are not calculated as the negative gradient of a potential energy function, but rather directly predicted by the model, without going through the potential energy. These forces are generally faster to compute than forces derived from the potential energy by automatic differentiation. However, these predictions must be used with care, see https://arxiv.org/abs/2412.11569.

In metatomic models, they are associated with the "non_conservative_force" or "non_conservative_force/<variant>" name (see Variants), and must have the following metadata:

Metadata for "non_conservative_force"

Metadata	Names	Description
keys	"_"	the keys must have a single dimension named "_", with a single entry set to 0. The "non_conservative_force" quantity is always represented as a metatensor.torch.TensorMap with a single block.
samples	["system", "atom"]	the samples must be named ["system", "atom"], since "non_conservative_force" is always per-atom. "system" must range from 0 to the number of systems given as input to the model. "atom" must range between 0 and the number of atoms/particles in the corresponding system. If selected_atoms is provided, then only the selected atoms for each system should be part of the samples.
components	"xyz"	The "non_conservative_force" quantity must have a single component dimension named "xyz", with three entries set to 0, 1, and 2. The non-conservative force is always a 3D vector, and the order of the components is x, y, z.
properties	"non_conservative_force"	The "non_conservative_force" quantity must have a single property dimension named "non_conservative_force", with a single entry set to 0.

The following simulation engines can use the "non_conservative_force" output, typically using a non_conservative flag:

ASE

i-PI

LAMMPS

Note

If you are adding support for non_conservative_force in a molecular dynamics engine, metatomic models might predict a non zero total force. You should consider removing this total force to prevent drift in your simulations.

Non-conservative stress

Similar to the "non_conservative_force", the non-conservative stress is a stress tensor that is not calculated using derivatives of the potential energy, but directly predicted by the model. As with forces, they are typically faster to compute but need to be used with care, see https://arxiv.org/abs/2412.11569.

In metatomic models, they are associated with the "non_conservative_stress" or "non_conservative_stress/<variant>" name (see Variants), and must have the following metadata:

Metadata for "non_conservative_stress"

Metadata	Names	Description
keys	"_"	the keys must have a single dimension named "_", with a single entry set to 0. The "non_conservative_force" quantity is always represented as a metatensor.torch.TensorMap with a single block.
samples	"system"	the samples must be named ["system"], since "non_conservative_stress" is always per-system. "system" must range from 0 to the number of systems given as input to the model.
components	["xyz_1"], ["xyz_2"]	the "non_conservative_stress" quantity must have two components labels with "xyz_1" and "xyz_2" as their respective names, both with three entries set to 0, 1, and 2. The order of the components along both directions is x, y, z.
properties	"non_conservative_stress"	the "non_conservative_stress" quantity must have a single property dimension named "non_conservative_stress", with a single entry set to 0.

The following simulation engines can use the "non_conservative_stress" output, typically using a non_conservative flag:

ASE

i-PI

LAMMPS