Skip to content

Presets

Available Presets#

Several presets are available, each with adjustable parameters:

  • FS: Closest to the classical PACE model, this preset uses linear radial functions and a Finnis-Sinclair-like embedding (read-out) function, similar to PACE, along with chemical embedding, which allows for an unlimited number of elements to be parameterized by the model. It can be exported and executed with a custom C++ implementation and supports MPI parallelization with domain decomposition in LAMMPS. Extrapolation grade based on D-optimality criteria can be used, after construction of active set.

  • GRACE_1LAYER: A one-layer model (i.e., only star graphs basis functions) extending the FS preset with MLP for radial functions and readouts, and additional improvements. This model supports only TensorFlow-based execution; GPU usage is highly recommended but not mandatory. Due to the local interactions (within a single cutoff distance), no extra overhead is required for domain decomposition, and the model can be parallelized with MPI in LAMMPS on multiple GPUs.

  • GRACE_2LAYER: A two-layer model (i.e., star and tree graph basis functions) that builds upon GRACE_1LAYER by adding a second layer of message passing. It also supports only TensorFlow-based execution, making GPU usage highly recommended.

  • Custom Model: Users can construct custom models by combining building blocks from the previous presets. More details on customization will be provided in future documentation.

Models hyperparameters#

Presets can be selected in input.yaml input file, in section potential:

potential:
  # LINEAR, FS, GRACE_1LAYER, GRACE_2LAYER
  preset: "FS" 

  # kw-arguments that will be passed to preset or custom function
  kwargs: {n_rad_base: 16, embedding_size: 64}

Latest and complete list of arguments can be found in definition of the models in the presets.py file.

FS#

Following parameters can be provided to tune FS model:

  • basis_type: "SBessel" - type of radial base functions: SBessel, Cheb
  • n_rad_base: 20 - number of radial base functions
  • embedding_size: 64 - number of chemical embedding channels
  • max_order: 4 - maximum product order for B-functions
  • n_rad_max: [20, 15, 10, 5] - max number of radial functions for each product order of output B-functions
  • lmax: [5, 5, 4, 3] - max l-character for each product order of input B-functions used to construct B-functions for current body order
  • Lmax: [None, 3, 0, 0] - max l-character for each product order of resulted output B-functions
  • max_sum_l: [None, None, 6, 4] - used to limit number of B-functions
  • fs_parameters: [[1.0, 1.0], [1.0, 0.5], [1.0, 2], [1.0, 0.75]] - parameters of FS embedding in a form [[c1,m1], [c2,m2], ...] that corresponds to \(c_1 \phi_1^{m_1} + c_2 \phi_2^{m_2} + ...\)

GRACE_1LAYER#

  • basis_type: "Cheb" - type of the radial basis functions: SBessel, Cheb
  • n_rad_base: 8 - number of radial basis functions
  • mlp_radial: True - use radial function as an MLP (True) or linear (False) transformation of the radial basis
  • embedding_size: 128 - size of the chemical embedding
  • max_order: 4 - maximum product order for B-functions
  • n_rad_max: 42 - max number of radial functions
  • lmax: 4 - maximum l-character for constructing product functions
  • n_mlp_dens: 16 - number of non-linear readout densities (+1 extra for linear readout automatically added)

GRACE_2LAYER#

  • basis_type: "Cheb" - type of radial basis functions: SBessel, Cheb
  • n_rad_base: 8 - number of radial basis functions
  • mlp_radial: True - use radial function as an MLP (True) or linear (False) transformation of the radial basis
  • embedding_size: 128 - size of the chemical embedding
  • max_order: 4 - maximum product order for B-functions
  • n_mlp_dens: 10 - number of non-linear readout densities (+1 extra for linear readout automatically added)
  • n_rad_max: [32, 48] - per-layer max number of radial functions
  • lmax: [3, 3] - maximum per-layer l-character for constructing product functions
  • indicator_lmax: 1 - maximum l-character of the first layer expansion (aka message) passed to the second layer