General information

Module for ab initio structure evolution (MAISE) is a C-based code for Linux platforms developed since 2009. It was originally designed as an evolutionary optimization engine interfaced with external density functional theory (DFT) packages to enable unconstrained ground state structure searches. The primary function of the present MAISE package is the construction of NN interatomic models for accurate mapping of ab initio potential energy surfaces.

Main features

Current MAISE version 2.3.13 (12 May 2021) features

  • neural network-based description of interatomic interactions

  • evolutionary optimization

  • structure analysis

1. The neural network (NN) module builds, tests, and uses NN models to describe interatomic interactions with near ab initio accuracy at a low computational cost compared to density functional theory calculations.

With the primary goal of using NN models to accelerate structure search, the main function of the module is to relax given structures. To simplify the NN application and comparison, we closely matched the input and output file formats with those used in the VASP software. Previously parameterized NN models available in the ‘models/’ directory have been generated and extensively tested for crystalline and/or nanostructured materials. First practical applications of NNs include the prediction of new synthesizable Mg-Ca alloys [1] and identification of more stable Cu-Pd-Ag nanoparticles [2].

Users can create their own NN models with MAISE which are typically trained on DFT total energy and atomic force data for relatively small structures. The generation of relevant and diverse configurations is done separately with an ‘evolutionary sampling’ protocol detailed in our published work [3]. The code introduces a unique feature, ‘stratified training’, of how to build robust NNs for chemical systems with several elements [3]. NN models are developed in a hierarchical fashion, first for elements, then for binaries, and so on, which enables generation of reusable libraries for extended blocks in the periodic table.

2. The evolutionary algorithm (EA) enables an efficient identification of ground state configurations at a given chemical composition. Our studies have shown that the EA is particularly advantageous in dealing with large structures when no experimental structural input is available [3], [4].

The searches can be performed for 3D bulk crystals, 2D films, and 0D nanoparticles. Population of structures can be generated either randomly or predefined based on prior information. Essential operations are ‘crossover’, when a new configuration is created based on two parent structures in the previous generation, and ‘mutation’, when a parent structure is randomly distorted. For 0D nanoparticles we have introduced a multitribe evolutionary algorithm that allows an efficient simultaneous optimization of clusters in a specified size range [2].

3. The analysis functions include the comparison of structures based on the radial distribution function (RDF), the determination of the space group and the Wyckoff positions with an external SPGLIB package, etc. In particular, the RDF-based structure dot product is essential for eliminating duplicate structures in EA searches and selecting different configurations in the pool of found low-energy structures.

An overview of the MAISE predictions and capabilities can be found in Ref. [5].

Inputs and libraries

Main input files that define a simulation are setup with job settings, model with NN parameters, and basis with the symmetry functions converting a structure into the NN input. The atomic structure is read from the POSCAR file that follows the VASP format.

EVOS

NNET

CELL

SEARCH

EXAM

PARSE

TRAIN

TEST

SIMUL

EXAM

setup

o

o

o

o

o

o

model

o !

o #

o #

basis

o

$

SPG

o

o

GSL

o

o
! for stratified training one needs to provide individual models
$ 'basis' stored in the parsed directory is appended to 'model' at the end of the training
# 'model' has 'basis' pasted at the end once training is finished

Units

All calculations within MAISE are done in electronvolts (eV) and Angstroms (A) for energies and distances, respectively.

The parameter units in the setup file are described in the corresponding sections, e.g., GPa for pressure, K for temperature, and seconds or femtoseconds for time.

The Behler-Parrinello symmetry functions for converting atomic environments into neural network inputs were originally given in Bohrs. Hence, the parameters in the basis file are usually defined in Bohrs but they can be specified in Angstroms as well by setting the conversion factor to 1.

Output energies, enthalpies, and volumes are given either per structure or per atom depending on the job type.

Resources

Latest information can be found on MAISE front page.

The source code, neural network models, and examples are available on Github repository.

The dev team welcomes questions, suggestions, and bug reports submitted to this Forum.

The package has been developed at the Univesity of Oxford (2009-2012) and Binghamton University (2012-present) by

_images/group.jpg

Alexey Kolmogorov

kolmogorov@binghamton.edu

Samad Hajinazar

hajinazar@binghamton.edu

Ernesto Sandoval

esandov1@binghamton.edu

Aidan Thorn

athorn1@binghamton.edu

Saba Kharabadze

skharab1@binghamton.edu

Funding support