Example Usage

Codes can be downloaded from the examples folder.

Usage in lammps is based on commands of the form fix <label> <atom_tag> mlmod <xml_parameter_file>.

Parameters

• <label>: name to give the instance of mlmod (multiple mlmod models can be used at the same time)

• <atom_tag>: type of atoms the mlmod model will be applied.

• <xml_parameter_file>: xml file specifying the mlmod model data.

For example, lines in a lammps script or in command calls are of the form:

fix ml_mobility1 all mlmod Main.mlmod_params

Parameter files <xml_parameter_file> are of the form:

<?xml version="1.0" encoding="UTF-8"?>
<MLMOD>
<model_data type="dX_MF_ML1">
<M_ii_filename value="M_ii_oseen1.pt"/>
<M_ij_filename value="M_ij_oseen1.pt"/>
</model_data> </MLMOD>

The type specifies the particular mlmod mode for modeling interactions, dynamics, or quantities-of-interest.

Exporting a PyTorch machine learning model to a .pt file is done using:

model = M_ii_Model(); # PyTorch ml model
x = torch.zeros((1,3));
traced_model = torch.jit.trace(model, (x));

torch_filename = 'M_ii_oseen1.pt';
traced_model.save(torch_filename);

For more details on generating the .pt machine learning models using PyTorch and for setting up full mlmod simulations, see the examples folder.

Below is a summary of the mlmod methods available for interfacing the machine learning approaches with simulations.

Name	Tag / Mode	Role	Description	MPI
General ML Force-Laws	F_ML1	force	General force law for groups of particles \(X=\{X_i\}_{i\in\mathcal{G}}\) based on a specified machine learning model, \(F(X,V,F,I_T,t)\).
General ML Dynamics	Dyn_ML1	dynamics	General dynamics based on specified machine learning model, \([X^{n+1},V^{n+1}] = \Gamma[X^n,V^n]\).
Quantity-of-Interest ML Methods	QoI_ML1	quantity of interest	For computing a Quantity-of-Interest (QoI) as defined by a specified machine learning model, \(A(X,V,F,I_T,t)\). Available for use in subsequent calculations within simulations.
General Mobility ML Model	dX_MF_ML1	dynamics	Mobility based on specified machine learning model, \({dX}/{dt} = M(X)F\).
Pairwise Mobility ML Model	dX_MF_Q1_ML1_Pair	dynamics	Mobility with pairwise-structure and specified machine learning model, \({dX}/{dt} = M(X)F + F_{thm}\).
Collective N-to-N Mobility ML Model	dX_MF_Q1_ML1_N2N	dynamics	Mobility with general structure and specified machine learning model, \({dX}/{dt} = M(X)F + F_{thm}\).
Mobility Test	dX_MF	dynamics	For testing mobility calculations, \({dX}/{dt} = M(X)F\).
Particle ML Force-Laws	F_X_ML1	force	Force law for individual particles \(X_i\) based on specified machine learning model, \(F_i(X_i,V_i,F_i,{I_T}_i,t)\).	yes
Pairwise ML Force-Law	F_Pair_ML1	force	Force law for pairs of particles based on specified machine learning model, \(F_{ij}(X_{ij},V_{ij},\) \(F_{ij},{I_T}_{ij},t)\).

[Examples Folder]