Guest editorial: Special Topic on software for atomistic machine learning Free

Welcome to the Journal of Chemical Physics’ Special Topic on Software for Atomistic Machine Learning. For some years now, search engines have been dominating our online experience and have essentially overtaken libraries, whether physical or digital, as the means to find information we are looking for. Most readers of an original research article find it by citation or direct search, and not by browsing journal volumes. Given this, one might wonder what the utility of a Special Topic issue of a scientific journal might be.

However, publishing papers on scientific software has traditionally been somewhat neglected, with few go-to journals for publishing, such as the Journal of Open Source Software or Computer Physics Communications. Typical published software papers tend to discuss relatively mature software packages. In this context, the Journal of Chemical Physics’ initiative¹ to support software publications is especially welcome. Given the huge activity currently taking place across many sub-fields and communities in new software development for atomistic machine learning (ML), this landscape is changing fast. Arguably, many papers in this Special Topic issue might not have been written if it were not for the impetus provided by this Special Topic issue.

Beyond their individual value regarding specific software packages, these papers as a collection provide a snapshot at this moment in time of the kinds of tools that people use and the goals they set themselves and achieve for the software implementations of their methods. Table I presents an overview of the 28 invited and contributed articles.^2–29 Of these, 18 (64%) deal directly with machine-learning interatomic potentials (MLIPs). The other ten articles cover a broad range of subjects, ranging from sampling to dataset repositories and workflows.

In the following, we give an overview of these contributions.

Since their beginnings in the 1980s and 1990s, MLIPs have undergone tremendous development and now constitute a highly active field of research. Some modern MLIPs can predict forces with an accuracy close to the underlying ab initio reference method for atomistic systems with many chemical elements and millions of atoms while still providing orders of magnitude of acceleration. These capabilities have increasingly enabled scientific applications using MLIPs that would not otherwise have been possible.

Consequently, there is a trend to directly integrate MLIPs into molecular dynamics codes. In this Special Topic, four contributions describe the integration of (a) neuro-evolution potentials into GPUMD (Graphics Processing Units Molecular Dynamics), including improved featurization, GPU code, active learning, and supporting Python packages gpyumd, calorine, and pynep;³ (b) PhysNet into CHARMM (Chemistry at HARvard Macromolecular Mechanics) via a new MLpot extension of the pyCHARMM interface, with para-chlorophenol as an example;¹² (c) general MLIPs into CASTEP (CAmbridge Serial Total Energy Package), including active learning, using the example of a Gaussian Approximation Potential (GAP)/smooth overlap of atomic positions (SOAP) model;¹⁶ and (d) ephemeral data-derived potentials with AIRSS (Ab Initio Random Structure Search).²⁵ 

An important milestone in MLIP development was the introduction of artificial neural network MLIPs that were able to efficiently handle large (“high-dimensional”) atomistic systems by Behler and Parrinello.³⁰ Six contributions in this Special Topic present MLIPs related to Behler–Parrinello networks: (a) neuro-evolution potentials;³ (b) the atomic energy network (ænet) in a PyTorch implementation for GPU support;⁸ (c) deep potentials via DeePMD-kit, recent improvements including attention-based features, learning dipoles and polarizabilities, long-range interactions, model compression, and GPU acceleration;¹⁸ (d) multi-layer-perceptron-based MLIPs via PANNA (Properties from Artificial Neural Network Architectures), with improved GPU support based on TensorFlow and long-range electrostatic interactions through a variational charge equilibration scheme;²¹ and (e) ephemeral data-derived potentials (EDDPs) for atomistic structure prediction, including ensemble-based uncertainties.²⁵ 

Message-passing neural networks allow the exchange of information between atoms beyond their local environments by repeatedly passing messages between them. Two contributions provide such MLIPs: (a) the SchNetPack2 library provides improved support functionality, including data sparsity, equivariance, and PyTorch-based MD, and provides four MLIPs: SchNet and FieldSchnet (external fields) and PaiNN and SO3net (equivariance);⁷ (b) the existing PhysNet MLIP is integrated into CHARMM.¹² 

Kernel-based learning is another ML approach that many MLIPs employ. Gaussian process regression (GPR), in particular, has been frequently used since the introduction of Gaussian Approximation Potentials (GAPs).³¹ Five contributions in this Special Topic deal with kernel-based MLIPs: (a) the quantum machine learning (QML)-lightning package provides GPU-accelerated sparse approximate GPR and representations (random features, FCHL19);⁴ (b) AL4GAP provides ensemble-based active learning for GAP MLIPs to study charge-neutral molten-salt mixtures,¹³ (c) the q-pac package implements kernel charge equilibration based on sparse GPR for long-range electrostatic interactions, non-local charge transfer, and energetic response to external fields;¹⁷ (d) the quantum mechanics and interatomic potentials (QUIP) package allows training and deployment of GAP models, recent additions including distributed training via the Message Passing Interface (MPI) and compressed features;²⁹ and (e) an application study develops a GAP/SOAP MLIP to obtain the pressure–temperature phase diagram of Pt and to simulate the spontaneous crystallization of a large Pt nanoparticle.⁶ 

Other ML approaches can be used to develop MLIPs, notably linear regression. Three contributions describe such MLIPs: (a) the ACEpotentials.jl Julia package provides MLIPs based on linear GPR and the atomic cluster expansion (ACE) representation, including uncertainties and active learning;²⁷ (b) the PESPIP (Potential Energy Surface Permutationally Invariant Polynomials) package provides MLIPs based on permutationally invariant polynomials in Morse-transformed interatomic distances, including their optimization;⁵ and (c) the MLIP-3 package provides moment tensor potentials, including fragment-based active learning for large simulation cells.²⁰ 

Besides the MLIPs themselves, seven contributions provide auxiliary tooling and analysis that focus on specific aspects of MLIPs but are not specific to one MLIP: (a) the DScribe library provides many atomistic representations and has been extended to include Valle–Oganov materials fingerprints and derivatives for all representations;⁹ (b) the sphericart package implements efficient real-valued spherical harmonics, a key ingredient of many representations for MLIPs, including stable Cartesian derivatives;¹⁹ (c) the XPOT (Cross-Platform Optimizer for Potentials) package provides hyperparameter optimization for MLIPs;¹⁴ (d) PES-Learn benchmarks four approaches to train neural-network MLIPs on data with different levels of fidelity (e.g., low and high accuracy);¹⁵ (e) the wfl (Workflow) and ExPyRe (Execute Python Remotely) packages provide workflow management routines tailored for atomistic simulations and MLIP development;²⁴ (f) the ColabFit Exchange repository hosts hundreds of diverse datasets of atomistic systems in extended XYZ format for MLIP benchmarking and development;²⁶ and (g) the glp package demonstrates how to use automatic differentiation to efficiently obtain forces, stress, and heat flux for message-passing MLIPs.²⁸ 

The remaining five contributions span a wide range of other areas: (a) the mlcolvar library implements multiple dimensionality reduction methods to identify collective variables for analysis and enhanced sampling in MD simulations, including an interface to the PLUMED (PLUgin for MolEcular Dynamics) software;¹⁰ (b) the AGOX (Atomistic Global Optimization X) package enables developing global optimization algorithms for atomistic structure search, including random search, basin hopping, evolutionary algorithms, and global optimization with first-principles energy expressions (GOFEE);² (c) the same AGOX package also includes structure generation based on local optimization in “complementary energy” landscapes (oversmoothed MLIPs), favoring structures with fewer distinct local motifs;¹¹ (d) the DeepQMC package provides a framework for neural network-based variational quantum Monte Carlo methods, including PauliNet, FermiNet, and DeepErwin;²² and (e) the SISSO++ (Sure Independence Screening and Sparsifying Operator) software offers symbolic regression, including recent improvements in expression representation, support for units, nonlinear parametrization, and the solver algorithm.²³ 

This Special Topic on Software for Atomistic Machine Learning contains 28 invited and contributed articles. They range from MLIPs based on neural networks, kernel models, and linear regression, as well as their integration into MD codes, to auxiliary tooling, structure search, dimensionality reduction, quantum Monte Carlo methods, and symbolic regression. We hope you enjoy reading this community effort at capturing the state of the field at this moment.

The Journal of Chemical Physics encourages and welcomes submissions of original articles describing software implementations relevant to the broad remit of the journal.

a-C

amorphous carbon

AD

automatic differentiation

ACE

atomic cluster expansion

AGOX

Atomistic Global Optimization X

AIRSS

Ab Initio Random Structure Search

AL

active learning

ANN

artificial neural network

ASE

Atomic Simulation Environment

ASL

Academic Software License

BO

Bayesian optimization

BSD

Berkeley Software Distribution

CAL

CASTEP Academic License

CASTEP

CAmbridge Serial Total Energy Package

CHARMM

Chemistry at HARvard Macromolecular Mechanics

CUDA

Compute Unified Device Architecture

DR

dimensionality reduction

EDDP

ephemeral data-derived potential

FCHL19

Faber, Christensen, Huang, Lilienfeld 2019

GAP

Gaussian approximation potential

GOFEE

global optimization with first-principles energy expressions

GPL

General Public License

GPR

Gaussian process regression

GPU

graphics processing unit

GPUMD

Graphics Processing Units Molecular Dynamics

JAX

Just After eXecution

LGPL

Lesser General Public License

LR

linear regression

MIT

Massachusetts Institute of Technology

MD

molecular dynamics

ML

machine learning

MLIP

machine-learning interatomic potential

MPI

message-passing interface

NLR

non-linear regression

PaiNN

Polarizable Atom Interaction Neural Network

PANNA

Properties from Artificial Neural Network Architectures

PES

potential energy surface

PIP

permutationally invariant polynomial

PLUMED

PLUgin for Molecular Dynamics

QML

quantum machine learning

QUIP

QUantum mechanics and Interatomic Potentials

RFF

random Fourier features

SISSO

Sure Independence Screening and Sparsifying Operator

SOAP

smooth overlap of atomic positions

SORF

structured orthogonal randomized features

SR

symbolic regression

XPOT

Cross-Platform Optimizer for Potentials

a-C

amorphous carbon

AD

automatic differentiation

ACE

atomic cluster expansion

AGOX

Atomistic Global Optimization X

AIRSS

Ab Initio Random Structure Search

AL

active learning

ANN

artificial neural network

ASE

Atomic Simulation Environment

ASL

Academic Software License

BO

Bayesian optimization

BSD

Berkeley Software Distribution

CAL

CASTEP Academic License

CASTEP

CAmbridge Serial Total Energy Package

CHARMM

Chemistry at HARvard Macromolecular Mechanics

CUDA

Compute Unified Device Architecture

DR

dimensionality reduction

EDDP

ephemeral data-derived potential

FCHL19

Faber, Christensen, Huang, Lilienfeld 2019

GAP

Gaussian approximation potential

GOFEE

global optimization with first-principles energy expressions

GPL

General Public License

GPR

Gaussian process regression

GPU

graphics processing unit

GPUMD

Graphics Processing Units Molecular Dynamics

JAX

Just After eXecution

LGPL

Lesser General Public License

LR

linear regression

MIT

Massachusetts Institute of Technology

MD

molecular dynamics

ML

machine learning

MLIP

machine-learning interatomic potential

MPI

message-passing interface

NLR

non-linear regression

PaiNN

Polarizable Atom Interaction Neural Network

PANNA

Properties from Artificial Neural Network Architectures

PES

potential energy surface

PIP

permutationally invariant polynomial

PLUMED

PLUgin for Molecular Dynamics

QML

quantum machine learning

QUIP

QUantum mechanics and Interatomic Potentials

RFF

random Fourier features

SISSO

Sure Independence Screening and Sparsifying Operator

SOAP

smooth overlap of atomic positions

SORF

structured orthogonal randomized features

SR

symbolic regression

XPOT

Cross-Platform Optimizer for Potentials