Bypassing the Kohn-Sham equations with machine learning.

Brockherde F, Vogt L, Li L, Tuckerman ME, Burke K, Müller KR. · 2017

Read Bypassing the Kohn-Sham equations with machine learning. by Brockherde F, Vogt L, Li L, Tuckerman ME, Burke K, Müller KR. (2017) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5636838.txt.

Introduction

Brockherde and co-authors report a methods study in Nature Communications. It tackles a costly step in quantum chemistry. Density functional theory often solves the Kohn-Sham equations. At least 30,000 papers used that scheme in one year. The team instead taught machines to learn the needed maps. Their method learns links between potential, density, and energy.

Earlier attempts stalled on a noisy derivative. Learning the Hohenberg-Kohn map directly sidesteps that problem. The authors test the maps on model systems and molecules. They then ran a molecular dynamics simulation with the learned functional. The test molecule was malonaldehyde. The run captured proton transfer inside the molecule. The authors say realistic density functionals can now be built this way.

Catalog record

CREATORBrockherde F, Vogt L, Li L, Tuckerman ME, Burke K, Müller KR.
YEAR2017
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5636838

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bypassing the Kohn-Sham equations with machine learning By Felix Brockherde, Leslie Vogt, Li Li, Mark E. Tuckerman, Kieron Burke, Klaus-Robert Müller Published in: Nature Communications DOI: 10.1038/s41467-017-00839-3 ABSTRACT Last year, at least 30,000 scientific papers used the Kohn–Sham scheme of density functional theory to solve electronic structure problems in a wide variety of scientific fields. Machine learning holds the promise of learning the energy functional via examples, bypassing the need to solve the Kohn–Sham equations. This should yield substantial savings in computer time, allowing larger systems and/or longer time-scales to be tackled, but attempts to machine-learn this functional have been limited by the need to find its derivative. The present work overcomes this difficulty by directly learning the density-potential and energy-density maps for test systems and various molecules. We perform the first molecular dynamics simulation with a machine-learned density functional on malonaldehyde and are able to capture the intramolecular proton transfer process. Learning density models now allows the construction of accurate density functionals for realistic molecular systems. INTRODUCTION Kohn–Sham density functional theory1 (KS-DFT) is now enormously popular as an electronic structure method in a wide variety of fields2. Useful accuracy is achieved with standard exchange–correlation (XC) approximations, such as generalized gradient approximations3 and hybrids4. Such calculations are playing a key role in the materials genome initiative5, at least for weakly correlated materials6. There has also been a recent spike of interest in applying machine learning (ML) methods in the physical sciences7–11. The majority of these applications involve predicting properties of molecules or materials from large databases of KS-DFT calculations12–15. A few applications involve finding potential energy surfaces within molecular dynamics (MD) simulations16–19. Fewer still have focussed on finding the functionals of DFT as a method of performing KS electronic structure calculations without solving the KS equations20–23,24. If such attempts could be made practical, the possible speed-up in repeated DFT calculations of similar species, such as occur in ab initio MD simulations, is enormous. A key difficulty has been the need to extract the functional derivative of the non-interacting kinetic energy. The non-interacting kinetic energy functional T s[n] of the electron density n is used in two distinct ways in a KS calculation1, as illustrated in Fig. 1: (i) its functional derivative is used in the Euler equation which is solved in the self-consistent cycle and (ii) when self-consistency is reached, the ground-state energy of the system is calculated by E[n], an orbital-free (OF) mapping. The solution of the KS equations performs both tasks exactly. Early results on simple model systems showed that ML could provide highly accurate values for T s[n] with only modest amounts of training20, but that the corresponding functional derivatives are too noisy to yield sufficiently accurate results to (i). Subsequent schemes overcome this difficulty in various ways, but they typically lose a factor of 10 or more in accuracy22, and their computational cost can increase dramatically with system complexity.Fig. 1Overview of approach and motivation. a Mappings used in this paper. The bottom arrow represents E[v], a conventional electronic structure calculation, i.e., KS-DFT. The ground-state energy is found by solving KS equations given the external potential, v. E[n] is the total energy density functional. The red arrow is the HK map n[v] from external potential to its ground state density. b (Top) How the energy error depends on M, the number of training points, for ML-OF and ML-HK with different basis sets for the 1D problem. b (Bottom) Errors in the PBE energies (relative to exact values) and the ML maps (relative to PBE) as a function of interatomic spacing, R, for H2 with M = 7. c How our Machine Learning Hohenberg–Kohn (ML-HK) map makes predictions. The molecular geometry is represented by Gaussians; many independent Kernel ridge regression models predict each basis coefficient of the density. We analyze the performance of data-driven (ML) and common physical basis representations for the electron density Here we present an alternative ML approach, in which we replaced the Euler equation by directly learning the Hohenberg–Kohn (HK) map v(r) → n(r) (red line in Fig. 1a) from the one-body potential of the system of interest to the interacting ground-state density, i.e., we establish an ML-HK map. We show that this map can be learned at a much more modest cost than either previous ML approaches to find the functional and its derivative (ML-OF) or direct attempts to model the energy as a functional of v(r) (ML-KS). Furthermore, we show that it can immediately be applied to molecular calculations by calculating the energies of small molecules over a range of conformers. Moreover, as we have already implemented this approach with a standard quantum chemical code (Quantum Espresso25) using a standard DFT approximation (Perdew–Burke–Ernzerhof (PBE)), this can now be tried on much larger scales. The ML-HK map reflects the underlying computational approach used to generate a particular electron density, but is not restricted to any given electronic structure method. Many molecular properties, including but not limited to the energy, are dependent on the electron density, making the ML-HK map more versatile than a direct ML-KS mapping. We also establish that densities can be learned with sufficient accuracy to distinguish between different DFT functionals, thus providing a route to future functional development via generation of precise densities for a range of molecules and conformations. RESULTS We will first outline theoretical results, most prominently the ML-HK map, and then illustrate the approach with simulations of one-dimensional (1D) systems and three-dimensional (3D) molecules. ML-HK map Previous results show that for an ML-OF approach, the accuracy of ML-KS kinetic energy models \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{\rm{s}}^{{\rm{ML}}}\left[ n \right]$$\end{document}TsMLn improve rapidly with the amount of data. However, minimizing the total energy via gradient descent requires the calculation of the gradient of the kinetic energy model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_{\rm{s}}^{{\rm{ML}}}$$\end{document}TsML (Fig. 1) and calculating this gradient is challenging. Due to the data-driven nature of, e.g., kernel models, the machine-learned kinetic energy functional has no information in directions that point outside the data manifold26. This heavily influences the gradient to an extent that it becomes unusable without further processing20. There have been several suggestions to remedy this problem, but all of them share a significant loss in accuracy compared to T s[n]21, 22, 27. Here, we propose an interesting alternative to gradients and the ML-OF approach. Recently, it has been shown that the HK map for the density as a functional of the potential can be approximated extremely accurately using semiclassical expressions28. Such expressions do not require the solution of any differential equation, and become more accurate as the number of particles increases. Errors can be negligible even for just two distinct occupied orbitals. Inspired by this success, we suggest how one might circumvent the kinetic energy gradient and directly train a multivariate ML model. We name this the ML-HK map:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${n^{{\rm{ML}}}}[v](x) = \mathop {\sum}\limits_{i = 1}^M {\beta _i}(x)k\left( {v,{v_i}} \right).$$\end{document}nML[v](x)= ∑i=1Mβi(x)kv,vi.Here each density grid point is associated with a group of model weights β. Training requires solving an optimization problem for each density grid point. Although this is possible in 1D, it rapidly becomes intractable in 3D, as the number

Read the complete text