Potential Model

Atomic dress

The absolute value of total energy from ab-initio calculations can often be very large. To avoid numerical problems, it is common practice to assign a constant atomic energy (dress) to each type of atom, such that the average energy is shifted to zero. Such an atomic dress can be generated with pinn.utils.get_atomic_dress

Loss function

The loss function in potential model is defined as following:

\[ L = w_e \cdot \mathrm{MSE}(e) + w_f \cdot \mathrm{MSE}(f) + w_s \cdot \mathrm{MSE} (s) + \mathrm{regularization} \]

Loss terms are mean squared errors of:

\(e\): energies
\(f\): forces components
\(s\): stress tensor components

It is assumed that the energy, force and stress labels are labelled as "e_data", "f_data", "s_data" in the training set respectively, if they are to be used in the loss function.

Parameters

Below is a list of additional parameters of the potential model and their descriptions.

Parameter	Default	Description
`e_dress`	`{}`	Atomic Dress
`e_scale`	`1`	The energy scaling during training, this variable defines the energy unit during training
`e_unit`	`1`	Unit of the input energies, this variable will be used in the ASE calculator
`max_energy`	`False`	When set to a number, exclude energies above this value in loss function
`use_e_per_atom`	`False`	Use the per-atom energy in place of total energy in loss function
`log_e_per_atom`	`True`	Log the per-atom energy error, automatically enabled with `use_e_per_atoms=True`
`use_e_weight`	`False`	Scale the energy loss according to the `'e_weight'` Tensor in the dataset
`use_force`	`False`	Include force in loss function
`max_force_comp`	`False`	if set to float, omit forces components larger than it
`no_force_comp`	`False`	if set to int, use as maximum number of force component per weight update
`use_f_weights`	`False`	Scale the energy loss according to the `'f_weights'` Tensor in the dataset
`use_stress`	`False`	Include stress in loss function
`use_l2`	`False`	Include L2 regularization in loss function
`e_loss_multiplier`	`1`	Weight of energy loss
`f_loss_multiplier`	`1`	Weight of force loss
`s_loss_multiplier`	`1`	Weight of stress loss
`l2_loss_multiplier`	`1`	Weight of l2 loss

ASE calculator

A calculator can be created from a model as simple as:

from pinn import get_calc
calc = get_calc('/path/to/model/')
calc = get_calc('/path/to/model/', default_dtype='float64')
calc.calculate(atoms)
calc.get_forces()

ASE MD is always float64 (positions, cell, momenta, the integrator).

default_dtype is a constructor argument only (not YAML), same idea as MACE model.float() / model.double(). None follows settings.dtype. If it differs, checkpoint weights are cast once at load and the TF graph runs in that dtype. See Models.

Units

Following the convention of ASE, the output unit is eV for energy, eV/Å for forces and eV/Å³ for stress tensor. Since PiNN does not know about the unit in the dataset, a to_eV parameter is given to convert from the dataset unit to eV (this optional is not necessary if e_unit is set in the parameters).

Available results

energy: total energy
forces: forces
energies: atomic contribution to the energy
stress: stress tensor