- Access throughyour institution
- View Open Manuscript
- Section snippets
- References (42)
- Cited by (133)
- Recommended articles (6)
Computer Physics Communications
Author links open overlay panelYuzhiZhangabHaidiWangcWeijieChendJinzheZengeLinfengZhangfPersonEnvelopeHanWanggPersonEnvelopeWeinanEafPersonEnvelope
In recent years, promising deep learning based interatomic potential energy surface (PES) models have been proposed that can potentially allow us to perform molecular dynamics simulations for large scale systems with quantum accuracy. However, making these models truly reliable and practically useful is still a very non-trivial task. A key component in this task is the generation of datasets used in model training. In this paper, we introduce the Deep Potential GENerator (DP-GEN), an open-source software platform that implements the recently proposed ”on-the-fly” learning procedure (Zhang et al. 2019) and is capable of generating uniformly accurate deep learning based PES models in a way that minimizes human intervention and the computational cost for data generation and model training. DP-GEN automatically and iteratively performs three steps: exploration, labeling, and training. It supports various popular packages for these three steps: LAMMPS for exploration, Quantum Espresso, VASP, CP2K, etc. for labeling, and DeePMD-kit for training. It also allows automatic job submission and result collection on different types of machines, such as high performance clusters and cloud machines, and is adaptive to different job management tools, including Slurm, PBS, and LSF. As a concrete example, we illustrate the details of the process for generating a general-purpose PES model for Cu using DP-GEN.
Program Title: DP-GEN
Program Files doi: http://dx.doi.org/10.17632/sxybkgc5xc.1
Licensing provisions: LGPL
Programming language: Python
Nature of problem: Generating reliable deep learning based potential energy models with minimal human intervention and computational cost.
Solution method: The concurrent learning scheme is implemented. Supports for sampling configuration space with LAMMPS, generating ab initio data with Quantum Espresso, VASP, CP2K and training potential models with DeePMD-kit are provided. Supports for different machines including workstations, high performance clusters and cloud machines are provided. Supports for job management tools including Slurm, PBS, LSF are provided.
In recent years, machine learning (ML) has emerged as a promising tool for the field of molecular modeling. In particular, ML-based models have been proposed to address a long-standing issue, the accuracy-vs-efficiency dilemma when one evaluates the potential energy surface (PES), a function of atomic positions and their chemical species, and its negative gradients with respect to the atomic positions, namely the interatomic forces. From a first-principles point of view, PES is derived from the many-particle Schrödinger equation under the Born–Oppenheimer approximation, and the interatomic forces are given naturally by the Hellman–Feynman theorem. To this end, the ab initio molecular dynamics (AIMD) scheme, wherein accurate PES and interatomic forces are obtained within the density functional theory (DFT) approximation, has been most widely adopted, , . Unfortunately, the cost of AIMD restricts its typical applications to system sizes of hundreds of atoms and the time scale of 100ps. In the opposite direction, efficient empirical PES models, or force fields (FF), allow us to perform much larger and longer simulations; but their accuracy and transferability is often an issue. ML has the potential to change this situation: a good ML model trained on ab initio data should have an efficiency comparable with that of FF models in the sense that the costs of ML models and FF models both scale linearly with system size, while maintaining ab initio accuracy.
Developing ML-based PES models involves two components, data generation and model construction. To date, most discussions have focused on the second component. Two important issues are: A good functional form (e.g.kernel based models or neural networks) and respecting physical constraints of the PES, such as the extensiveness and symmetry properties. In this regard, two representative classes of models have emerged: the kernel-based models like the Gaussian Approximation Potential and the neural network (DNN) based models like the Behler–Parrinello model and the Deep Potential model, . In particular, the smooth version of the Deep Potential model is an end-to-end model that satisfies the requirements mentioned above.
There have also been some efforts on open-source software along this line, , , . Of particular relevance to this work is the DeePMD-kit package, which has been developed to minimize the effort required to build DNN-based PES models and to perform DNN-based MD simulation. DeePMD-kit is interfaced with TensorFlow, one of the most popular deep learning frameworks, making the training process highly automatic and efficient. DeePMD-kit is also interfaced with popular MD packages, such as the LAMMPS package for classical MD and the i-PI package for path integral MD. Thus, once one has a good sets of data, there are now effective tools for training Deep Potentials that can be readily used to perform efficient molecular dynamics simulation for all different kinds of purposes.
In comparison, much less effort has gone into the first component mentioned above: data generation. In spite of the tremendous interest and activity, very few have taken the effort to make sure that the dataset used to train the ML-based PES is truly representative enough. Indeed data generation is often quite ad hoc. Some notable exceptions are found in, , . In Ref., an active learning procedure was proposed based on an existing unlabeled dataset. Some data points in that set are selected to be labeled, and the result is then used to train the ML model. The procedure ensures that the selected dataset is at least representative of the original unlabeled dataset. In Refs., , one begins with no data, labeled or unlabeled, and explores the configuration spaces following some systematic procedure. For each of the configurations encountered, a decision is made as to whether that configuration should be labeled. The exploration procedure is designed to ensure that the unlabeled dataset, i.e.all the configurations encountered in the exploration, is representative of all the situations that the ML-based model is intended for. Even though these procedures were also described as “active learning”, there is a difference since in these procedures, one does not have the unlabeled data to begin with and choosing the unlabeled data judiciously is also an important part of the whole procedure.
To highlight this difference, we will call the procedures in Refs.,  “concurrent learning”. By “concurrent learning”, we mean that one does not have any data to begin with, labeled or unlabeled, and the data is generated on the fly as the training proceeds. The generation of the data and the learning is an interactive process to ensure that one obtains an “optimal” dataset which is on one hand representative enough and on the other hand as small as possible. This is in contrast to “sequential learning” in which the data is generated beforehand and the training of the model is performed afterwards. It also differs from active learning in the sense that active learning starts with unlabeled data. In a way the purpose of active learning is to find the smallest dataset that needs to be labeled in an existing unlabeled dataset.
The actual concurrent learning procedure goes as follows. One uses different sampling techniques (such as direct MD at different thermodynamic conditions, enhanced sampling, Monte Carlo) based on the current approximation of the PES to explore the configuration space. An efficient error indicator (this is the error in the PES) is then used to monitor the snapshots generated during the sampling process. Those that have significant errors will then be selected and sent to a labeling procedure, in which accurate ab initio energies and forces are calculated and added to the training dataset. A new approximation of the PES is obtained by training with the accumulated training dataset. These steps are repeated until convergence is achieved, i.e.,the configuration space has been explored sufficiently, and a representative set of data points has been accurately labeled. At the end of this procedure, a uniformly accurate PES model is generated. We refer to Ref. for more details.
In order to carry out such a procedure efficiently, one also needs reasonable computational resources. Since many tasks can be done in parallel, one needs to implement automatic and efficient parallel processing algorithms. Taking the exploration stage for example, it may happen that dozens to hundreds of MD simulations are executed simultaneously with different initial configurations under different thermodynamic conditions. If these tasks are executed manually, it will require a great deal of human labor, not to mention the compromise in efficiency. It would be even worse if one wants to utilize different computational resources in different concurrent learning steps, e.g.,a high performance cluster (HPC) with most advanced GPU nodes for training and an HPC with a vast number of CPU nodes for labeling. Selecting a machine with the most available computational power among a group of candidate machines is also an issue. For all these reasons, we feel that it would be useful for the molecular and materials simulation community to have an open-source implementation of the concurrent learning procedure which, among other things, can automatically schedule the iterative process, dispatch different computational tasks to different computational resources, and collect and analyze the results.
In this paper, we introduce DP-GEN, an open-source concurrent learning platform and software package for the generation of reliable deep learning based PES models, in a way that minimizes the computational cost and human intervention. We describe the implementation of DP-GEN, which is based on the procedure proposed in Ref.. We will focus on two modules, the scheduler and the task dispatcher. A modularized coding structure for the scheduler is designed, making it possible to incorporate different methods or software packages for the three different components in the concurrent learning procedure: exploration, labeling, and training. The dispatcher module is prepared for handling a huge number of tasks in a high-throughput fashion, and it is made compatible with different kinds of machines and popular job scheduling systems.
This paper is organized as follows. In Section2 we present the basic methodology that the DP-GEN workflow follows. In Section3 we introduce the details of the software, including how the concurrent learning process is scheduled and how different tasks are dispatched. In Section4, we give a concrete example, in which a general purpose Deep Potential model for Cu is generated using DP-GEN. Conclusions and outlooks are given in the last Section.
The DP-GEN workflow contains a series of successive iterations. Each iteration is composed of three steps: exploration, labeling, and training. We denote by , abbreviated , the PES represented by the DP model, where denotes atomic positions and denotes the parameters. An important point throughout the DP-GEN procedure is that we have an ensemble of models trained from the same set of data but with difference in the initialization of model parameters . evolves
Implemented with Python, DP-GEN provides a user-friendly interface. The master process can be started via a single line of command: where the arguments PARAM and MACHINE are both the names of parameter files in the json format that specify the user’s demands.
DP-GEN is composed of two major modules. First, DP-GEN serves as a scheduler, which follows the aforementioned concurrent learning scheme and generates computational tasks iteratively for the three steps: exploration, labeling and
In this section, we report the details of the process that we follow to generate a DP model for Cu with DP-GEN, and demonstrate its uniform accuracy when used to predict a broad range of properties.
In this paper, we introduced the software platform DP-GEN. We described its implementation and reported the details when used to generate a general purpose PES model for Cu. We expect DP-GEN to be a scalable and flexible platform. The three steps, exploration, training, and labeling, which are controlled by the scheduler, are separate and highly modularized. Therefore, developers will spend a minimal amount of effort to incorporate novel functionalities. For example, DP-GEN can easily be
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
The authors thank Marcos F. Calegari Andrade, Hsin-Yu Ko, Jianxing Huang, Yunpei Liu, Mengchao Shi, Fengbo Yuan, and Yongbin Zhuang for helps and discussions. We are grateful for computing time provided by the TIGRESS High Performance Computer Center at Princeton University, the High-performance Computing Platform of Peking University, and the Beijing Institute of Big Data Research. The work of L. Z. and W. E was supported in part by a gift from iFlytek to Princeton University, USA, the ONR, USA
- PizziG. et al.
Comput. Mater. Sci.
- TogoA. et al.
- KresseG. et al.
Comput. Mater. Sci.
- PodryabinkinE.V. et al.(Video) Keynote: Scalable and Distributed DNN Training on Modern HPC Systems
Comput. Mater. Sci.
- CeriottiM. et al.
Comput. Phys. Comm.
J. Comput. Phys.
- WangH. et al.
Comput. Phys. Comm.
- ArtrithN. et al.
Comput. Mater. Sci.
- KohnW. et al.
- CarR. et al.
Phys. Rev. Lett.
Ab Initio Molecular Dynamics: Basic Theory and Advanced Methods
Phys. Rev. Lett.
Phys. Rev. Lett.
Commun. Comput. Phys.
Phys. Rev. Lett.
J. Chem. Theory Comput.
Tensorflow: Large-scale machine learning on heterogeneous distributed systems
J. Chem. Phys.
Phys. Rev. Mater.
- Lattice dynamics and elastic properties of α-U at high-temperature and high-pressure by machine learning potential simulations
2022, Journal of Nuclear Materials
Studying the physical properties of materials under high pressure and temperature through experiments is difficult. Theoretical simulations can compensate for this deficiency. Currently, large-scale simulations using machine learning force fields are gaining popularity. As an important nuclear energy material, the evolution of the physical properties of uranium under extreme conditions is still unclear. Herein, we trained an accurate machine learning force field on α-U and predicted the lattice dynamics and elastic properties at high pressures and temperatures. The force field agrees well with the ab initio molecular dynamics (AIMD) and experimental results and it exhibits higher accuracy than classical potentials. Based on the high-temperature lattice dynamics study, we first present the temperature-pressure range in which the Kohn anomalous behavior of the Σ4 optical mode exists. Phonon spectral function analysis showed that the phonon anharmonicity of α-U is very weak. We predict that the single-crystal elastic constants C44, C55, C66, polycrystalline modulus (E, G), and polycrystalline sound velocity (CL, CS) have strong heating-induced softening. All the elastic moduli exhibited compression-induced hardening behavior. The Poisson's ratio shows that it is difficult to compress α-U at high pressures and temperatures. Moreover, we observed that the material becomes substantially more anisotropic at high pressures and temperatures. The accurate predictions of α-U demonstrate the reliability of the method. This versatile method facilitates the study of other complex metallic materials.(Video) Success in the Age of Digital X The Quadrant of IoT Scale Databases
- Molecular dynamics simulations of LiCl ion pairs in high temperature aqueous solutions by deep learning potential
2022, Journal of Molecular Liquids
Molecular dynamics simulation is an efficient method to study ion-pair association in high temperature supercritical fluid. Interatomic potentials based on neural-network machine learning shows outstanding ability of balancing the accuracy and the efficiency in molecular dynamics (MD) simulations. In present study, a neural-network potential (NNP) model for LiCl ion-pair in high temperature aqueous solutions was developed using database obtained by the first-principles density functional theory (DFT) calculations. With this NNP model, the structures of LiCl solution and the dissociation pathway of LiCl dissociation process were investigated. The results show that deep learning molecular dynamic (DPMD) simulations can accurately reproduce the radial distribution functions of ab initio MD simulations. And several metastable states were clearly identified from the dissociation 2D energy surfaces. In addition, the potential of mean force (PMF) profiles and corresponding association constants (Ka) were extensively investigated under a wide range of temperature-density (T=330–1273K and ρ=0.45–1.0g/cm3) conditions. The association constants calculated from DPMD are satisfactory compared with the experimental data. The study indicates that deep learning potential exhibits good capabilities to describe the association behavior of metal complex in high temperature aqueous solutions. This work also provides the microstructures and LiCl association constants in temperature aqueous solutions for which no experimental data exist.
- Toward full ab initio modeling of soot formation in a nanoreactor
A neural network (NN)-based model is proposed to construct the potential energy surface of soot formation. Our NN-based model is proven to possess good scalability of O(N) and retain the ab initio accuracy, which allows the investigation of the entire evolution of soot particles with tens of nm from an atomic perspective. A series of NN-based molecular dynamics (NNMD) simulations are performed using a nanoreactor scheme to investigate the critical process in soot formation – the inception of PAH radicals. The results show that physical interaction enhances chemical inception, and such enhancement is observed for clusters of π- and σ-radicals, which are distinct from the dimer. We also observed that PAH radicals of ∼400Da can produce core-shell soot particles at a flame temperature, with a disordered core and outer shell of stacked PAHs, suggesting a potential physically stabilized soot inception mechanism.
- High accuracy neural network interatomic potential for NiTi shape memory alloy
2022, Acta Materialia
Nickel-titanium (NiTi) shape memory alloys (SMA) are widely used, however simulating the martensitic transformation of NiTi from first principles remains challenging. In this work, we developed a neural network interatomic potential (NNIP) for near-equiatomic Ni-Ti system through active-learning based acquisitions of density functional theory (DFT) training data, which achieves state-of-the-art accuracy. Phonon dispersion and potential-of-mean-force calculations of the temperature-dependent free energy have been carried out. This NNIP predicts temperature-induced, stress-induced, and defect-induced martensitic transformations from atomic simulations, in significant agreement with experiments. The NNIP can directly simulate the superelasticity of NiTi nanowires, providing a tool to guide their design.
- Interatomic potentials for oxide glasses: Past, present, and future
2022, Journal of Non-Crystalline Solids: X
The continuous development and improvement of interatomic potential models for oxide glasses have made classical molecular dynamics a powerful computational technique routinely used for studying the structure and properties of such materials on a par with the more advanced experimental techniques.
In this brief review, we retrace the development of the most used interatomic potential models from the earliest MD simulations up to now with a look at the possible future developments in this field due to the advent of the machine learning era and data-driven methods.
- The chemical origin of temperature-dependent lithium-ion concerted diffusion in sulfide solid electrolyte Li<inf>10</inf>GeP<inf>2</inf>S<inf>12</inf>
2022, Journal of Energy Chemistry
Solid-state batteries have received increasing attention in scientific and industrial communities, which benefits from the intrinsically safe solid electrolytes (SEs). Although much effort has been devoted to designing SEs with high ionic conductivities, it is extremely difficult to fully understand the ionic diffusion mechanisms in SEs through conventional experimental and theoretical methods. Herein, the temperature-dependent concerted diffusion mechanism of ions in SEs is explored through machine-learning molecular dynamics, taking Li10GeP2S12 as a prototype. Weaker diffusion anisotropy, more disordered Li distributions, and shorter residence time are observed at a higher temperature. Arrhenius-type temperature dependence is maintained within a wide temperature range, which is attributed to the linear temperature dependence of jump frequencies of various concerted diffusion modes. These results provide a theoretical framework to understand the ionic diffusion mechanisms in SEs and deepen the understanding of the chemical origin of temperature-dependent concerted diffusions in SEs.
Research articleDeep Density: Circumventing the Kohn-Sham equations via symmetry preserving neural networks
Journal of Computational Physics, Volume 443, 2021, Article 110523
The recently developed Deep Potential [Phys. Rev. Lett. 120 (2018) 143001 ] is a powerful method to represent general inter-atomic potentials using deep neural networks. The success of Deep Potential rests on the proper treatment of locality and symmetry properties of each component of the network. In this paper, we leverage its network structure to effectively represent the mapping from the atomic configuration to the electron density in Kohn-Sham density function theory (KS-DFT). By directly targeting at the self-consistent electron density, we demonstrate that the adapted network architecture, called the Deep Density, can effectively represent the self-consistent electron density as the linear combination of contributions from many local clusters. The network is constructed to satisfy the translation, rotation, and permutation symmetries, and is designed to be transferable to different system sizes. We demonstrate that using a relatively small number of training snapshots, with each snapshot containing a modest amount of data-points, Deep Density achieves excellent performance for one-dimensional insulating and metallic systems, as well as systems with mixed insulating and metallic characters. We also demonstrate its performance for real three-dimensional systems, including small organic molecules, as well as extended systems such as water (up to 512 molecules) and aluminum (up to 256 atoms).(Video) IBM and ASTRON Unveil 64-bit Microserver Prototype
Research articleA DFT accurate machine learning description of molten ZnCl2 and its mixtures: 1. Potential development and properties prediction of molten ZnCl2
Computational Materials Science, Volume 185, 2020, Article 109955
Molten eutectic salts consisting of ZnCl2 and other alkali chlorides are promising thermal storage and heat transfer fluid materials in the next generation concentrated solar thermal power. To go deep into the thermal and transport properties for a high order mixture, the microstructure information, as well as thermodynamics properties of individual components, have to be identified first. This work develops interatomic potentials of molten ZnCl2 based on neural-network machine learning approach for the first time. The machine learning potential is trained by fitting to the energies and forces of liquid structures ab initio molecular dynamics calculations. The developed machine learning potential is validated by comparing partial radial distribution functions, coordination numbers, and partial structure factors with AIMD and PIM potential. The machine learning potential yields a more precise description of the microstructures than the PIM potential which suffers from the analytical form. Furthermore, structural and thermophysical evolution with temperature are studied and the results are in good agreement with experimental values. The efficient machine learning potential with DFT accuracy from our study will provide a promising scheme for accurate molecular simulations of structures and dynamics of molten ZnCl2 mixtures.
Research articleShort- and medium-range orders in Al90Tb10 glass and their relation to the structures of competing crystalline phases
Acta Materialia, Volume 204, 2021, Article 116513
Molecular dynamics simulations using an interatomic potential developed by artificial neural network deep machine learning are performed to study the local structural order in Al90Tb10 metallic glass. We show that more than 80% of the Tb-centered clusters in Al90Tb10 glass have short-range order (SRO) with their 17 first coordination shell atoms stacked in a ‘3661’ or ‘15551’ sequence. Medium-range order (MRO) in Bergman-type packing extended out to the second and third coordination shells is also clearly observed. Analysis of the network formed by the ‘3661’ and ‘15551’ clusters show that ~82% of such SRO units share their faces or vertexes, while only ~6% of neighboring SRO pairs are interpenetrating. Such a network topology is consistent with the Bergman-type MRO around the Tb-centers. Moreover, crystal structure searches using genetic algorithm and the neural network interatomic potential reveal several low-energy metastable crystalline structures in the composition range close to Al90Tb10. Some of these crystalline structures have the ‘3661’ SRO while others have the ‘15551’ SRO. While the crystalline structures with the ‘3661’ SRO also exhibit the MRO very similar to that observed in the glass, the ones with the ‘15551’ SRO have very different atomic packing in the second and third shells around the Tb centers from that of the Bergman-type MRO observed in the glassy phase.
Research articleActive learning of linearly parametrized interatomic potentials
Computational Materials Science, Volume 140, 2017, pp. 171-180
This paper introduces an active learning approach to the fitting of machine learning interatomic potentials. Our approach is based on the D-optimality criterion for selecting atomic configurations on which the potential is fitted. It is shown that the proposed active learning approach is highly efficient in training potentials on the fly, ensuring that no extrapolation is attempted and leading to a completely reliable atomistic simulation without any significant decrease in accuracy. We apply our approach to molecular dynamics and structure relaxation, and we argue that it can be applied, in principle, to any other type of atomistic simulation. The software, test cases, and examples of usage are published at http://gitlab.skoltech.ru/shapeev/mlip/.
Research articleSpectral neighbor analysis method for automated generation of quantum-accurate interatomic potentials
Journal of Computational Physics, Volume 285, 2015, pp. 316-330
We present a new interatomic potential for solids and liquids called Spectral Neighbor Analysis Potential (SNAP). The SNAP potential has a very general form and uses machine-learning techniques to reproduce the energies, forces, and stress tensors of a large set of small configurations of atoms, which are obtained using high-accuracy quantum electronic structure (QM) calculations. The local environment of each atom is characterized by a set of bispectrum components of the local neighbor density projected onto a basis of hyperspherical harmonics in four dimensions. The bispectrum components are the same bond-orientational order parameters employed by the GAP potential . The SNAP potential, unlike GAP, assumes a linear relationship between atom energy and bispectrum components. The linear SNAP coefficients are determined using weighted least-squares linear regression against the full QM training set. This allows the SNAP potential to be fit in a robust, automated manner to large QM data sets using many bispectrum components. The calculation of the bispectrum components and the SNAP potential are implemented in the LAMMPS parallel molecular dynamics code. We demonstrate that a previously unnoticed symmetry property can be exploited to reduce the computational cost of the force calculations by more than one order of magnitude. We present results for a SNAP potential for tantalum, showing that it accurately reproduces a range of commonly calculated properties of both the crystalline solid and the liquid phases. In addition, unlike simpler existing potentials, SNAP correctly predicts the energy barrier for screw dislocation migration in BCC tantalum.
Research articleTheoretical prediction on the local structure and transport properties of molten alkali chlorides by deep potentials
Journal of Materials Science & Technology, Volume 75, 2021, pp. 78-85
In this work, the local structure and transport properties of three typical alkali chlorides (LiCl, NaCl, and KCl) were investigated by our newly trained deep potentials (DPs). We extracted datasets from ab initio molecular dynamics (AIMD) calculations and used these to train and validate the DPs. Large-scale and long-time molecular dynamics simulations were performed over a wider range of temperatures than AIMD to confirm the reliability and generality of the DPs. We demonstrated that the generated DPs can serve as a powerful tool for simulating alkali chlorides; the DPs also provide results with accuracy that is comparable to that of AIMD and efficiency that is similar to that of empirical potentials. The partial radial distribution functions and angle distribution functions predicted using the DPs are in close agreement with those derived from AIMD. The estimated densities, self-diffusion coefficients, shear viscosities, and electrical conductivities also matched well with the AIMD and experimental data. This work provides confidence that DPs can be used to explore other systems, including mixtures of chlorides or entirely different salts.
This paper and its associated computer program are available via the Computer Physics Communication homepage on ScienceDirect (http://www.sciencedirect.com/science/journal/00104655)(Video) Next@Acer 2019 | Live from New York
The review of this paper was arranged by Prof. Stephan Fritzsche.
© 2020 Elsevier B.V. All rights reserved.