Machine learning unifies the modeling of materials and molecules

Bartók, Albert P.; De, Sandip; Poelking, Carl; Bernstein, Noam; Kermode, James R.; Csänyi, Gábor; Ceriotti, Michele

doi:10.1126/sciadv.1701816

Cited by 648 publications

(718 citation statements)

References 63 publications

Supporting

Mentioning

710

Contrasting

Order By: Relevance

“…[104] Looking toward the future, it seems promising to fit to such RPA data (given that gradients and thus force data are available), rather than DFT. 2007 First neural-network (NN) potential for bulk silicon and tests for the liquid phase Behler and Parrinello [40] 2008 Description of high-pressure phase transitions in silicon using NN potentials and metadynamics Behler et al [92,93] 2010 First Gaussian approximation potential (GAP) model for diamond-type silicon and tests for elastic properties Bartók et al [68] 2015 Development of an on-the-fly learning scheme using the example of silicon Li et al [80] 2017 Correct energy ranking of the silicon (111) and (100) surface reconstructions using a GAP Bartók et al [39] 2018 Description of surface energies, dislocations, cracks, etc. a) Representative structures from a large reference database, for which DFT energies have been computed.…”

Section: High Accuracy For Crystalline and Amorphous Materials: The Cmentioning

confidence: 99%

Machine Learning Interatomic Potentials as Emerging Tools for Materials Science

2019

Self Cite

View full text Add to dashboard Cite

In pursuit of this goal, atomic-scale computer simulations have long been a central approach, and two major families of methods are routinely used today. On the one hand, there are quantum-mechanical simulations, in which we solve Schrödinger's equation for the electronic structure of molecular and periodic systems, most widely based on density-functional theory (DFT). [6][7][8] These methods provide (largely) reliable results for structural models of materials that normally contain a few tens or hundreds of atoms. State-of-the-art DFT methods can be applied to many material classes, and they are increasingly used for high-throughput screening and "in silico" (computer-based) design of materials: new compositions and previously unknown structures have been identified in DFT searches and subsequently experimentally realized. [5,[9][10][11] On the other hand, interatomic potential models ("force fields"), parameterizing interactions between atoms with (relatively) simple functional forms, are widely used in materials science to describe matter in molecular dynamics (MD) simulations. These simulations grant access to larger time and length scales, reaching system sizes of up to hundreds of thousands of atoms. [12] In parameterizing these potentials, a certain physical form of the atomic interactions is assumed, often in terms of bond distances, angles, and so on, and physical properties such as equilibrium lattice parameters or elastic constants enter the fitting of the potential. For this reason, such potentials are often called "empirical." They are several orders of magnitude faster than DFT, but necessarily less accurate and less easily transferable.In this Progress Report, we highlight recent developments in "machine-learned" interatomic potentials, which represent a rapidly growing field that promises to do away with the aforementioned trade-offs. Over the last year, there has been a surge of interest in machine learning (ML) methodology: part of it is due to the dramatic growth of ML throughout the scientific disciplines, and part of it is due to tangible success stories of ML-based interatomic potentials that are now beginning to emerge. We will argue that this is an exciting development with very practical implications, currently on the verge of moving from a somewhat specialized new technology to everyday applicability, poised to enhance and complement the communities' existing strengths in computational materials modeling. We will show selected applications of ML potentials to problems in materials science, discuss the current limitations (and possible pitfalls), and outline what we expect to be interesting directions for the development of the field in the coming years.Atomic-scale modeling and understanding of materials have made remarkable progress, but they are still fundamentally limited by the large computational cost of explicit electronic-structure methods such as density-functional theory. This Progress Report shows how machine learning (ML) is currently enabling a new degree of realism in materi...

show abstract

Section: High Accuracy For Crystalline and Amorphous Materials: The Cmentioning

confidence: 99%

Machine Learning Interatomic Potentials as Emerging Tools for Materials Science

2019

Self Cite

View full text Add to dashboard Cite

show abstract

“…51) and has since been used to generate potentials for diverse molecular and solid-state materials. 40,[52][53][54] A high-dimensional t to reference energy and force data is performed based on structural similarity or kernel functions, comparing atomic environments one by one. The initial choice for these have been many-body descriptors, most importantly the Smooth Overlap of Atomic Positions (SOAP), which includes all neighbours of an atom up to a cut-off radius.…”

Section: Gap Ttingmentioning

confidence: 99%

Data-driven learning and prediction of inorganic crystal structures

et al. 2018

Self Cite

View full text Add to dashboard Cite

Crystal structure prediction algorithms, including ab initio random structure searching (AIRSS), are intrinsically limited by the huge computational cost of the underlying quantum-mechanical methods. We have recently shown that a novel class of machine learning (ML) based interatomic potentials can provide a way out: by performing a highdimensional fit to the ab initio energy landscape, these potentials reach comparable accuracy but are orders of magnitude faster. In this paper, we develop our approach, dubbed Gaussian approximation potential-based random structure searching (GAP-RSS), towards a more general tool for exploring configuration spaces and predicting structures. We present a GAP-RSS interatomic potential model for elemental phosphorus, which identifies and correctly "learns" the orthorhombic black phosphorus (A17) structure without prior knowledge of any crystalline allotropes. Using the tubular structure of fibrous phosphorus as an example, we then discuss the limits of free searching, and discuss a possible way forward that combines a recently proposed fragment analysis with GAP-RSS. Examples of possible tubular (1D) and extended (3D) hypothetical allotropes of phosphorus as found by GAP-RSS are discussed. We believe that in the future, ML potentials could become versatile and routine computational tools for materials discovery and design.

show abstract

“…An indispensable ingredient to most machine learning models are molecular descriptors, which are constructed to provide an invariant, unique and efficient representation as input to machine learning models [18][19][20][21][22][23][24]. A popular molecular descriptor is the bag-of-bonds (BOB) model [25], which is an extension of the Coulomb matrix (CM) approach [4] and groups the pairwise distances according to pairs of atom types.…”

Section: Introductionmentioning

confidence: 99%

Capturing intensive and extensive DFT/TDDFT molecular properties with machine learning

et al. 2018

View full text Add to dashboard Cite

Abstract. Machine learning has been successfully applied to the prediction of chemical properties of small organic molecules such as energies or polarizabilities. Compared to these properties, the electronic excitation energies pose a much more challenging learning problem. Here, we examine the applicability of two existing machine learning methodologies to the prediction of excitation energies from time-dependent density functional theory. To this end, we systematically study the performance of various 2-and 3-body descriptors as well as the deep neural network SchNet to predict extensive as well as intensive properties such as the transition energies from the ground state to the first and second excited state. As perhaps expected current state-of-the-art machine learning techniques are more suited to predict extensive as opposed to intensive quantities. We speculate on the need to develop global descriptors that can describe both extensive and intensive properties on equal footing.

show abstract

Machine learning unifies the modeling of materials and molecules

Abstract: Statistical learning based on a local representation of atomic structures provides a universal model of chemical stability.

Cited by 648 publications

References 63 publications

Machine Learning Interatomic Potentials as Emerging Tools for Materials Science

Machine Learning Interatomic Potentials as Emerging Tools for Materials Science

Data-driven learning and prediction of inorganic crystal structures

Capturing intensive and extensive DFT/TDDFT molecular properties with machine learning

Contact Info

Product

Resources

About