Psi-k - Ab initio (from electronic structure) calculation of complex processes in materials


Reactive classical potentials versus hybrid methods: toward chemical complexity


TITLE: Reactive classical potentials versus hybrid methods: toward chemical complexity

Principal Organizer:

Jean-Bernard Maillet
Commissariat a l'Energie Atomique
Bruyeres-le-Chatel, FRANCE
Phone: +33 1 69 26 73 36
Fax: +33 1 69 26 70 77
Email:

Co-organizers:

(1) Timothy C.Germann
Los Alamos National Laboratory
X-7, MailStop D413, Los Alamos,
NM 87545, U.S.A.
Phone (505) 665-9772
Fax (505) 667-3726
Email

(2) Alejandro Strachan
California Institute of Technology
Materials and Process Simulation Center
Beckman Institute (139-74)
California Institute of Technology
Pasadena, CA 91125, U.S.A.
Phone (626) 395-8137
Fax (626) 585-0918
Email

(3) Alessandro De Vita
University of Trieste
Materials Engineering Dept.
via A. Valerio 2
I-34127 Trieste, Italy
Phone +39 040 558 3704
Fax +39 040 572 044
Email

(4) Michael C. Payne
University of Cambridge
Cavendish Laboratory (TCM)
University of Cambridge
Madingley Rd, Cambridge CB3 OHE, UK
Phone +44 1223 337381
Fax +44 1223 337356
Email


Schedule

We propose formal talks Monday through Thursday, closed by a working group session on Thursday afternoon. Anytime in June or August 2003 would be preferred. A few participants will be attending a conference "APS on Shock Compression of Condensed Matter", Portland, Oregon, USA on July 20-25, so the weeks immediately preceding or following this may be inconvenient.


Budget

The budget is based on 36 participants, 30 of which (25 speakers and 5 organizers) will request per diem support for 4 days. In euro:

4 nights 6000
4 days 3600
Travel expenses for 9 2745
overseas participants
Workshop dinner + incidentals 1100

Total requested 13445

Besides financing from CECAM, the organizers plan to obtain funding from the Ψk ESF programme. These two sources of funding are expected to contribute in about equal parts. Travel expenses of one of the US participants will be sponsored by ACCELRYS. The travel and per diem expenses of the six non-presenting participants are assumed covered by funding from the EU Network "Complex Atom Modelling".


Format

We propose having approximately 25 speakers, not counting the 5 organizers. The formal part of the workshop is planned for four days, with 7 talks per day on Monday through Wednesday, and 4 talks on Thursday morning, concluding with a focused round-table discussion in the afternoon.

Proposal


1. Scientific background

Current ab initio methods, although very accurate and general, are computationally extremely demanding. Their application is thus restricted to relatively small systems and short simulation times. On the other hand, the use of classical force fields allow the dynamical simulation of millions of atoms, which makes them applicable to the study of a wide variety of physical processes (e.g., shockwaves,dislocation dynamics, fracture, oxidation) which require larger system sizes and longer simulation times. However, force fields are notoriously difficult to develop and their accuracy must be established for each application.

A main current challenge is to develop methodologies that retain the accuracy of quantum mechanics (QM) while allowing large-scale simulations. Current approaches to bridge from electronic structure to atomistics can be categorized in two groups, described in the next sections.

(I) Improvement of the techniques for force-field development. The scope is to achieve classical potentials capable of describing chemical reactions. This aims at the production of transferable reactive force fields, normally based on accurate QM calculations which sample the various chemical environments encountered in the applications.

(II) Development of embedding schemes. Here concurrent, multi-scale models are produced in which high quality QM calculations are performed in a small space-time region (where/when important chemistry is occurring) and a simpler (non reactive) force field description is used to simulate the dynamics of the rest of the system.

2.1 Reactive force fields

In general, the total energy of an atomistic system is written as the sum of three terms, E_elec, E_val, and E_vdW, where E_elec denotes electrostatic interactions, E_val denotes valence interactions (bonds, angles, torsions, etc.), and E_vdW denotes van der Waals interactions (short range repulsion and longer range attraction). In order to describe chemical reactions and atomic interactions in a changing chemical environment, most reactive force fields are based on one (or both) of two main concepts:

I. A dependence of E_val on the partial bond order, obtained solely from atomic positions.

II. Environment-dependent, self-consistent charges in E_elec (including atomic polarizability if necessary).

These variable charges and partial bond orders are the key concepts that allow a classical description of chemical reactions, dramatically improving the transferability of interatomic potentials (e.g. retaining the same description for oxygen atoms in O2, H2O and Al2O3). Concerning the derivation of the partial bond orders, reactive force fields can be categorized into two main groups: (i) Tight-binding derived analytic moment expansions [1]. (ii) Physically sound functionals parameterized using either high quality QM or experiments [2-4].

The two existing models of the latter class of reactive force fields are the REBO potential (and its further developments, AIREBO [4]) and the quantum-based force field denoted ReaxFF [2]. Both models were originally developed for hydrocarbons, but now include (or will include soon) the extension to other systems (nitrogen, oxides, and metals). In order to gain accuracy on the description of chemical reactions, these force fields have to be fitted on a large variety of QM calculations including bond dissociation curves, angular and torsional bending, and the decomposition mechanisms of various molecules. To enable this without unnecessary duplication of effort, the 2002 CECAM workshop community developed the idea of building a web site hosted by CECAM (currently under development) for such interatomic potential training sets.


2.2 Embedding schemes

Another general approach for "bridging the length scale gap" between the quantum and the classical world is based on embedding strategies in which the accuracy of the description is not uniform throughout the investigated system.

In the standard embedding approach the system subregion where high accuracy is needed is described by a QM Hamiltonian, and is surrounded by ("embedded" in) a larger zone treated by classical interatomic potentials. In some applications, a third level of description (e.g., a finite element description) is used to include another external comtinuum region [9]. The system is thus divided into different parts associated to different models (levels of sophistication), and the main difficulties are encountered in matching different models at the boundary regions. For instance, the so-called QM/MM (Quantum Mechanics/ Molecular Mechanics) methods find their applications in various fields, including biophysics and biochemistry, where the chemically interesting parts are treated with ab initio techniques, while the surrounding regions are treated by empirical potentials. Although allowing for "large-scale" simulations (several thousand atoms treated classically, plus typically a hundred treated by QM), these methods have about the same time limitations of the pure ab initio methods when applied on small system sizes. This could become a problem when the reaction time of the chemical processes of interest is much longer than the feasible simulation time. In these cases, it is necessary to develop special techniques for sampling rare events [5]. Another important open problem in these concurrent simulations is how to choose the regions to be treated by different models, and in particular how these regions should evolve during the course of the dynamics.


In a different spirit the LOTF (Learn On The Fly) technique uses a unique (parametrized, classical) Hamiltonian to describe the atomic interactions. However, when needed, the interatomic potential parameters associated to the "chemically active" atoms (e.g. atoms in a crack tip) are allowed to vary during the simulation to better describe their changing environments [6]. The continuously adjusted parameters, which have a local meaning, can be obtained by "learning" from QM calculations carried out "on the fly" on restricted sets of atoms relevant to the local chemical processes observed during the simulation, using e.g. a force matching technique. With this procedure, an arbitrarily accurate description of the dynamical evolution of a system can be achieved, computing QM information only when/where needed (time embedding is thus introduced, on the top of the more standard space embedding). While fitting the parameters "on the fly" introduces some extra complication in the algorithm, the use of a unique Hamiltonian effectively removes matching problems in the boundary regions. This method is similar in some respects to another approach based on classical potentials, the Optimal Potential method developed by A. Laio and S. Bernard [7] for metals, in which accuracy is given the priority over transferability. Here the interaction potential is characterized by an enlarged set of parameters (not resolved in space) which are fitted on first principles calculations on the entire system. As the first principles calculations can be substantially separated in time, the time limitations of this method are essentially those of a classical approach.

Most of these hybrid length-scale methods encounter problems concerning the conservation of energy, since the Hamiltonian is changed during the simulation when boundary regions move or potential parameters are changed, and none address the equally important time-scale problems [5]. An original alternative in this direction, coupling in time and using parallel computers, has been introduced by G. Zerah [8]. The basic idea of this method is that a high accuracy dynamical trajectory (e.g. ab initio MD) can be obtained in a computationally efficient way starting from an approximate trajectory (e.g. using an empirical potential) and making a series of iterative corrections that require independent dynamical runs using the high-accuracy method but for times much smaller than the total simulation time. If these corrections are made in parallel, the target accurate trajectory can be obtained in significantly less wall-clock time (but more CPU time) than a complete high-accuracy dynamical run would require. One mathematical property of the relevant equations is that the convergence of this process is guaranteed. Multi-scale modeling in material science is a natural application of such a method (where the fine and coarse grain propagators are ab initio and classical descriptions respectively). This method is currently being developed and a wide range of applications is envisaged.


3. Motivation and objectives

The proposed workshop is a natural continuation of the CECAM-SIMU workshop held in July 2002 at CECAM on "Upscaling from ab initio to Molecular Dynamics: Interatomic Potentials and Hybrid Methods" (J.-B. Maillet, T.C. Germann and A. Strachan) with a change of focus based on the discussions and conclusions of the previous workshop. The workshop in 2002 was concerned with accurate atomistic simulations of various materials (metal alloys and oxides, organics, high-energy materials, and biological systems) using reactive force fields and hybrid embedding schemes.


It was clear that a great deal of effort is being devoted to developing the computational tools to study complex materials and processes (complex chemistry and phase transitions, large length and time scale simulations, etc.) and important progress has been made in various areas. It was also clear that: (i) further developments in the atomistic methodologies are required in order to solve outstanding problems in science and technology. These problems (in "real" materials, i.e. requiring low-symmetry simulated systems of nanometric scale or bigger) require more dedicated effort in the direction of handling chemical complexity. Moreover, (ii) further progress toward the validation of atomistic simulations against accurate experimental data is necessary and(iii) further progress toward linking atomistic simulations to larger-scale models (mesoscopic or macroscopic) needs to be made in order to simulate real conditions for many applications.

Consequently we propose to focus the 2003 workshop on developing reactive (bond order) potentials and multi-scale concurrent (embedding) methods which can be applied to complex materials and processes of scientific or technological importance}. We will concentrate on:

(1) Development of atomistic simulation methods, both using classical reactive force fields and multi-scale concurrent (embedding) techniques;

(2) Validation of the atomistic simulations. It is of great importance to critically assess the accuracy, range of application and predictive power of the current state of the art methods

(3) Use of the atomistic information to obtain insight and parameters that characterize the fundamental processes that govern the macroscopic behavior of materials, and their use with larger-scale models (coarse-grain force fields, kinetic Monte Carlo,accelerated dynamics, phase fields, micro-mechanical modeling, etc.) to achieve length and time scales of real materials and processes;

(4) Case-study applications (already achieved, or perspective) of these techniques in a number of specific "real" problems. Computer simulation practitioners with a strong experimental background will be invited to attend the workshop and present detailed insight into specific "challenging" problems in need of a suitable modeling technique.


In summary, the proposed workshop will bring together world experts in various field such as Physics, Chemistry, Materials Science and Engineering to address critical issues regarding the use of ab initio quantum mechanical methods and data and classical interatomic potentials to perform large-scale simulations of a variety of materials. We will focus on recent advances that allow the study of complex phenomena such as condensed phase chemical reactions.

Since this area of research is in constant development, a key objective of the proposed workshop is to gather and discuss the latest results obtained on complex systems. Also, we plan to take advantage of the workshop to address the status of the web-site that came out of the 2002 workshop, and its possible evolutions (possible extensions include codes, simulation results, and preprints). We also expect that due to the intense current activity and of the many problem remaining open in this field, several collaborations will be initiated or strengthened during this workshop.

Tentative list of participants.

References

  1. D.G. Pettifor, I.I.Oleinik, Phys. Rev. Lett.85, 4136 (2000).
  2. A.C.T. van Duin, S. Dasgupta, F. Lorant, and W.A.Goddard III,
    J.Phys.Chem.A 105, 9396 (2001).
  3. A.Strachan, T.Cagin, O.Gulseren, S.Mukherjee,
    R.E.Cohen, and W.A.Goddard III, Phys.Rev.B (submitted).
  4. S.J.Stuart, A.B.Tutein, and J.A.Harrison,
    J.Chem.Phys.112, 6472 (2000).
  5. A.F.Voter, F.Montalenti, and T.C.Germann,
    Annu.Rev. Mater.Sci.32, 321 (2002).
  6. A.De Vita and R.Car, Mat.Res.Soc.Symp.Proc.491, 473 (1998).
  7. A.Laio, S.Bernard, G.L.Chiarotti, S.Scandolo, and
    E. Tosatti, Science 287, 1027 (2000).
  8. L.Baffico, S.Bernard, Y. Maday, G.Turinici and G.
    Zerah , accepted for publication in Phys. Rev. E.
  9. F.F.Abraham, J.Q.Broughton, N.Bernstein and E.Kaxiras,
    Comput. Physics 12, 538 (1998).