Machine Learning Interatomic Potentials¶

exaStamp ships several machine-learning interatomic potentials (MLIPs). Each one lives in its own plugin under src/potential/mlip-<name> and is disabled by default: turn it on at configure time with the matching CMake option (see Build & Install).

Model	Page	Force operator	Init operator	CMake option	External dependency
SNAP	SNAP	`snap_force_fp32`, `snap_force_fp64`	`snap_init` (descriptors only)	`EXASTAMP_MLIP_SNAP_BUILD`	exaNBody built with `EXANB_BUILD_CONTRIB_MD=ON`
POD	POD	`pod_force`	`pod_init`	`EXASTAMP_MLIP_POD_BUILD`	BLAS / LAPACK
MTP	MTP	`mtp_force`	`mtp_init`	`EXASTAMP_MLIP_MTP_BUILD`	none
NNP	NNP (n2p2)	`n2p2_force`	—	`EXASTAMP_MLIP_N2P2_BUILD`	n2p2 installation
k2b	k2b	`k2b_compute_force`, `k2b_multi_force`, ...	`k2b_init`	always built	none

Init / force pattern¶

Most MLIPs split their setup from the force evaluation:

<model>_init goes in the init_parameters block. It reads the potential files once, before setup_system, and raises rcut_max to the potential cutoff. This way the ghost layers and neighbor lists are sized correctly from the start, and you don't need to set rcut_max by hand under global.
<model>_force goes in the compute_force block and evaluates energies, forces and the virial at every step.

Usage example

init_parameters:
  - species
  - pod_init:
      parameters:
        pod_file: "Ta_param.pod"
        coeff_file: "Ta_coefficients.pod"

compute_force:
  - pod_force

Ghost contributions

MLIP force operators accumulate forces on ghost atoms. Forces and energies must be zeroed in the ghosts before the force computation and folded back into their owner atoms afterwards. Include the shipped configuration file:

includes:
  - config_update_symmetric_forces.msp

or, equivalently, write the two blocks yourself:

compute_force_prolog:
  - zero_force_energy: { ghost: true }
compute_force_epilog:
  - update_force_energy_from_ghost
  - force_to_accel

Species mapping

How the species of the species block are matched to the elements of the potential files depends on the model. SNAP matches them by position: declare the species in the same order as the element blocks of the coefficient file. PACE matches them by name. Check the page of each model.

Descriptors and training¶

SNAP, POD, MTP and k2b can also compute their descriptors without evaluating the potential: per-atom descriptors, their derivatives, and the global design matrix (energy, forces and virial rows) used to fit linear models. This includes a loop over a database of configurations. See Descriptors.