4.7 Article

Accelerated path integral methods for atomistic simulations at ultra-low temperatures

期刊

JOURNAL OF CHEMICAL PHYSICS
卷 145, 期 5, 页码 -

出版社

AMER INST PHYSICS
DOI: 10.1063/1.4959602

关键词

-

资金

  1. Cluster of Excellence RESOLV - Deutsche Forschungsgemeinschaft [EXC 1069]
  2. Swiss National Science Foundation [200021-159896]
  3. Swiss National Science Foundation (SNF) [200021_159896] Funding Source: Swiss National Science Foundation (SNF)

向作者/读者索取更多资源

Path integral methods provide a rigorous and systematically convergent framework to include the quantum mechanical nature of atomic nuclei in the evaluation of the equilibrium properties of molecules, liquids, or solids at finite temperature. Such nuclear quantum effects are often significant for light nuclei already at room temperature, but become crucial at cryogenic temperatures such as those provided by superfluid helium as a solvent. Unfortunately, the cost of converged path integral simulations increases significantly upon lowering the temperature so that the computational burden of simulating matter at the typical superfluid helium temperatures becomes prohibitive. Here we investigate how accelerated path integral techniques based on colored noise generalized Langevin equations, in particular the so-called path integral generalized Langevin equation thermostat ( PIGLET) variant, perform in this extreme quantum regime using as an example the quasi-rigid methane molecule and its highly fluxional protonated cousin, CH5+. We show that the PIGLET technique gives a speedup of two orders of magnitude in the evaluation of structural observables and quantum kinetic energy at ultralow temperatures. Moreover, we computed the spatial spread of the quantum nuclei in CH4 to illustrate the limits of using such colored noise thermostats close to the many body quantum ground state. Published by AIP Publishing.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

4.7
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

Article Multidisciplinary Sciences

Simulating the ghost: quantum dynamics of the solvated electron

Jingang Lan, Venkat Kapil, Piero Gasparotto, Michele Ceriotti, Marcella Iannuzzi, Vladimir V. Rybkin

Summary: The nature of the bulk hydrated electron has been a challenge for both experiment and theory. Here the authors use a machine-learning model trained on MP2 data to achieve an accurate determination of the structure, diffusion mechanisms, and vibrational spectroscopy of the solvated electron.

NATURE COMMUNICATIONS (2021)

Article Materials Science, Multidisciplinary

Quantum vibronic effects on the electronic properties of solid and molecular carbon

Arpan Kundu, Marco Govoni, Han Yang, Michele Ceriotti, Francois Gygi, Giulia Galli

Summary: The study investigates the impact of quantum vibronic coupling on the electronic properties of carbon allotropes, utilizing path integral first principles molecular dynamics combined with a colored noise thermostat. By avoiding common approximations and only adding a moderate computational cost to FPMD simulations, the approach is suitable for large supercells needed for describing amorphous solids. The research predicts the effect of electron-phonon coupling on the fundamental gap of amorphous carbon and reveals a larger zero-phonon renormalization of the band gap in diamond than previously reported.

PHYSICAL REVIEW MATERIALS (2021)

Article Chemistry, Physical

Equivariant representations for molecular Hamiltonians and N-center atomic-scale properties

Jigyasa Nigam, Michael J. Willatt, Michele Ceriotti

Summary: This article discusses a family of structural descriptors that generalize atom-centered density correlation features to multiple centers and shows how they can be applied to learning matrix elements of the single-particle Hamiltonian. These features are fully equivariant in terms of translations, rotations, and permutations of indices associated with atoms.

JOURNAL OF CHEMICAL PHYSICS (2022)

Article Chemistry, Physical

Local Kernel Regression and Neural Network Approaches to the Conformational Landscapes of Oligopeptides

Raimon Fabregat, Alberto Fabrizio, Edgar A. Engel, Benjamin Meyer, Veronika Juraskova, Michele Ceriotti, Clemence Corminboeuf

Summary: The application of machine learning to theoretical chemistry has enabled the combination of quantum chemical energetics accuracy with thorough sampling of fluctuations. By utilizing local kernel regression and neural networks, accurate potential energy surfaces were generated efficiently to sample the conformational landscape of polypeptides.

JOURNAL OF CHEMICAL THEORY AND COMPUTATION (2022)

Article Chemistry, Physical

Unified theory of atom-centered representations and message-passing machine-learning schemes

Jigyasa Nigam, Sergey Pozdnyakov, Guillaume Fraux, Michele Ceriotti

Summary: Data-driven schemes provide a concise and effective description of the arrangement of atomic constituents in molecular and crystal structures. Atom-centered density correlations (ACDC) are used as a basis to expand targets in a body-ordered and symmetry-adapted manner. The ACDC framework can be extended to include multi-centered information and provide a linear basis for regressing symmetric functions of atomic coordinates, making it suitable for both atom-centered and message-passing machine-learning schemes.

JOURNAL OF CHEMICAL PHYSICS (2022)

Editorial Material Chemistry, Physical

Comment on Manifolds of quasi-constant SOAP and ACSF fingerprints and the resulting failure to machine learn four-body interactions [J. Chem. Phys. 156, 034302 (2022)]

Sergey N. N. Pozdnyakov, Michael J. J. Willatt, Albert P. P. Bartok, Christoph Ortner, Gabor Csanyi, Michele Ceriotti

Summary: This paper points out the degenerate pairs of configurations issue in all low-body-order atom-density correlation representations of molecular structures and discovers the quasi-constant smooth overlap of atomic position and atom-centered symmetry function fingerprint manifolds closely related to this problem. They demonstrate that the rigid singular configurations can only occur in finite, discrete sets and propose methods to optimize model parameters and the training set to mitigate their impact on machine learning models.

JOURNAL OF CHEMICAL PHYSICS (2022)

Article Chemistry, Physical

A Machine Learning Model of Chemical Shifts for Chemically and Structurally Diverse Molecular Solids

Manuel Cordova, Edgar A. Engel, Artur Stefaniuk, Federico Paruzzo, Albert Hofstetter, Michele Ceriotti, Lyndon Emsley

Summary: Nuclear magnetic resonance (NMR) chemical shifts provide information about the atomic environments in solid materials, but accurately predicting these shifts is computationally expensive. ShiftML is a machine-learning model that can rapidly predict chemical shifts with the same accuracy as density functional theory (DFT) calculations, reducing the computational cost by over four orders of magnitude.

JOURNAL OF PHYSICAL CHEMISTRY C (2022)

Article Chemistry, Physical

Thermodynamics and dielectric response of BaTiO3 by data-driven modeling

Lorenzo Gigli, Max Veit, Michele Kotiuga, Giovanni Pizzi, Nicola Marzari, Michele Ceriotti

Summary: The study develops an integrated machine learning model that accurately describes the structural, energetic, and functional properties of ferroelectric materials, and allows for the investigation of the microscopic mechanism of ferroelectric transition without the need for a coarse-grained description. The main driver of the ferroelectric transition in barium titanate is found to be the order-disorder transition of the off-centered states.

NPJ COMPUTATIONAL MATERIALS (2022)

Article Chemistry, Physical

A smooth basis for atomistic machine learning

Filippo Bigi, Kevin K. Huguenin-Dumittan, Michele Ceriotti, David E. Manolopoulos

Summary: Machine learning frameworks based on correlations of interatomic positions rely on a discretized description of atomic density, and this study investigates a basis obtained from solving the Laplacian eigenvalue problem within a sphere around the atom of interest. The results show that this basis has controllable smoothness and performs better than some widely used basis sets, comparable to data-driven bases.

JOURNAL OF CHEMICAL PHYSICS (2022)

Article Chemistry, Physical

Electronic-Structure Properties from Atom-Centered Predictions of the Electron Density

Andrea Grisafi, Alan M. Lewis, Mariana Rossi, Michele Ceriotti

Summary: The electron density of molecules or materials is an important target quantity for machine learning models. This study proposes a model that uses multicentered atomic basis to represent the scalar field, allowing transferable and linear-scaling predictions. The model overcomes the challenges posed by the non-orthogonality of the basis by using a gradient-based approach and optimizing the feature space. The results show that this method can accurately predict electron density with a comparatively small training set and achieve a good balance between accuracy and computational efficiency.

JOURNAL OF CHEMICAL THEORY AND COMPUTATION (2022)

Article Chemistry, Multidisciplinary

A data-driven interpretation of the stability of organic molecular crystals

Rose K. Cersonsky, Maria Pakhnova, Edgar A. Engel, Michele Ceriotti

Summary: We have automated the construction of an extensive library of molecular building blocks using machine learning techniques. By predicting the stability of crystal structures based on the inherent structure and energy correlations, as well as the contribution of chemical groups to the lattice energy, this library provides a comprehensive database for guiding the design of molecular materials.

CHEMICAL SCIENCE (2023)

Editorial Material Chemistry, Physical

2021 JCP Emerging Investigator Special Collection

Michele Ceriotti, Lasse Jensen, David E. Manolopoulos, Todd Martinez, David R. Reichman, Francesco Sciortino, C. David Sherrill, Qiang Shi, Carlos Vega, Lai-Sheng Wang, Emily A. Weiss, Xiaoyang Zhu, Jenny Stein, Tianquan Lian

JOURNAL OF CHEMICAL PHYSICS (2023)

Article Materials Science, Multidisciplinary

Beyond potentials: Integrated machine learning models for materials

Michele Ceriotti

Summary: Machine learning-based interatomic potentials have become an essential tool in atomic-scale modeling of materials, enabling accurate predictions of properties and extending simulation capabilities. These models have greatly facilitated the calculation of thermodynamics and functional properties, and are bringing us closer to predictive simulations of existing and novel materials.

MRS BULLETIN (2022)

Article Computer Science, Artificial Intelligence

Incompleteness of graph neural networks for points clouds in three dimensions

Sergey N. Pozdnyakov, Michele Ceriotti

Summary: Graph neural networks (GNN) are popular in machine learning and have been successful in predicting properties of molecules and materials. However, first-order GNNs are known to be incomplete, leading to the design of more complex schemes. The construction of graph representations for molecules adds a geometric dimension, with the most common approach being to consider atoms as vertices and connect them with bonds. This approach, known as distance graph NNs (dGNN), has shown excellent resolving power in chemical ML. However, the authors present a counterexample that proves dGNNs are not complete even for fully-connected graphs induced by 3D atom clouds.

MACHINE LEARNING-SCIENCE AND TECHNOLOGY (2022)

Article Materials Science, Multidisciplinary

Predicting hot-electron free energies from ground-state data

Chiheb Ben Mahmoud, Federico Grasselli, Michele Ceriotti

Summary: The article discusses the prediction of electronic free energy at arbitrary electron temperature using machine learning, avoiding the need to train temperature-dependent potentials, providing a new method. By combining physical considerations with machine-learning predictions, it offers a blueprint for the development of atomistic modeling.

PHYSICAL REVIEW B (2022)

暂无数据