Molecular dynamics (MD), empirical potentials, and molecular statics Kai Nordlund Professor, Department of Physics Adjoint Professor, Helsinki Institute of Physics Vice dean, Faculty of Science University of Helsinki, Finland Prof. Kai Nordlund Principal investigator Dr Carolina Björkas Fusion reactor mat'ls M Sc Laura Bukonte Fusion reactor mat’ls M Sc Elnaz Safi Fusion reactor mat’ls Doc. Antti Kuronen Principal investigator Dr Jussi Polvi Fusion reactor mat’ls M Sc Wei Ren Carbon nanostructures M Sc Alvaro Lopez Surface ripples Doc. Flyura Djurabekova Principal investigator Doc. Krister Henriksson Nuclear Materials Dr Andrea Sand Fusion reactor mat’ls Dr. Andrey Ilinov Nanomechanics Dr Vahur Zadin* Particle physics mat’ls (Univ. of Tartu) M Sc Fredric Granberg Dislocations M Sc Morten Nagel Nuclear materials M Sc Kostya Avchachov Dr. Aleksi Leino Irradiation of metals Nanostructures in silica M Sc Shuo Zhang Ion range calculations Kai Nordlund, Department of Physics, University of Helsinki Mr Jesper Byggmästar Nanowires M Sc Ekaterina Baibuz Particle physics mat'ls Dr Stefan Parviainen Particle physics mat'ls M Sc Mihkel Veske Particle physics mat'ls Dr Ville Jansson Particle physics mat'ls Dr Zhenxing Wang Particle physics mat'ls M Sc Junlei Zhao Nanoclusters M Sc Anders Korsbäck Particle physics mat’ls (CERN) M Sc Mrunal Parekh M Sc Simon Vigonski Particle physics mat'ls Particle physics mat'ls 2 Contents Molecular dynamics method General approach Nonequilibrium extensions -- briefly Interatomic potentials Overview And of EAM-like and Tersoff-like potentials the Albe philosophy to fit bond-order potentials Case study: Nanocluster simulations And introduction to tutorial case Molecular statics MD and conjugate gradient relaxation schemes The end: some practical advices Examples of recent applications Time permitting Kai Nordlund, Department of Physics, University of Helsinki 3 Introduction: The multiscale simulation framework m Length Finite Elements Rate equations mm Discrete dislocation dynamics μm nm DFT ps Classical Molecular dynamics ns Kinetic Monte Carlo μs ms s hours years Time Kai Nordlund, Department of Physics, University of Helsinki 4 MD method MD = Molecular dynamics MD is solving the Newton’s (or Lagrange or Hamilton) equations of motion to find the motion of a group of atoms Originally developed by Alder and Wainwright in 1957 to simulate atom vibrations in molecules Hence the name “molecular” Name unfortunate, as much of MD done nowadays does not include molecules at all First materials science application: 1960, Gibson simulated radiation effects in solids [Phys. Rev. 120 (1960) 1229)] A few hundred atoms, very primitive repulsive potentials (cell was actually under huge pressure) Kai Nordlund, Department of Physics, University of Helsinki 5 MD method Simple (trivial) example of atom motion by MD: thermal motion in Cu at 600 K, cross section of 2 atom layers Kai Nordlund, Department of Physics, University of Helsinki 6 MD method Less trivial example: small atom collision event 1 keV Xe irradiation of Cu Kai Nordlund, Department of Physics, University of Helsinki 7 MD method Nanocluster deposition event Kai Nordlund, Department of Physics, University of Helsinki 8 MD method How is this achieved: MD algorithm, the basic idea Give atoms initial r(t=0) and v(0) , choose short Δt Get forces F = − ∇ V(r(i)) or F = F(Ψ) and a = F/m Move atoms: r(i+1) = r(i) +v(i)Δt + 1/2 a Δt2 + correction terms Update velocities: v(i+1) = v(i) +aΔt + correction terms Move time forward: t = t + Δt Repeat as long as you need Kai Nordlund, Department of Physics, University of Helsinki 9 MD method MD algorithm, in greater detail Give atoms initial r(i=0) and v(i=0) , set a = 0.0, t = 0.0, i = 0, choose short Δt Predictor stage: predict next atom positions: Move atoms: rp = r(i) + v(i) Δt + 1/2 a Δ t2 + more accurate terms Update velocities: vp = v(i) + a Δt + more accurate terms Get forces F = - ∇ V(rp) or F = F( Ψ(rp) ) and a = F/m Corrector stage: adjust atom positions based on new a: Move atoms: r(i+1) = rp + some function of (a, Δt) Update velocities: v(i+1) = vp + some function of (a, Δt) Apply boundary conditions, temperature and pressure control as needed Calculate and output physical quantities of interest Move time and iteration step forward: t = t + Δt, i = i + 1 Repeat as long as you need Kai Nordlund, Department of Physics, University of Helsinki 10 MD method MD – atom representations MD naturally needs atom coordinates (and velocities) Atom coordinates can simply be read in from an ASCII text file Simple but for atoms good enough format: .XYZ Arrays in an MD code, e.g.: double precision :: x(MAXATOMS),y(MAXATOMS),z(MAXATOMS) Kai Nordlund, Department of Physics, University of Helsinki 11 MD method Initializing atom positions Atom coordinates can (should) be initialized to the structure of interest Simplest: single crystal: just make atom coordinates according to basis vectors. Simple 4D loop over (x,y,z,basis vectors) Some codes like LAMMPS have built-in construction routines Other possible cases are numerous Biomolecules: use databases of coordinates Amorphous material: very tricky to get something similar to experiments. Simplest: start from random atom coordinates, simulate at high T and cool down slowly. But this gives often too much coordination defects. Other approach: WootenWeaire-Winer Monte Carlo, can generate coordinationdefect free materials Quasicrystals: not known how to do it! Kai Nordlund, Department of Physics, University of Helsinki 12 MD method Initializing velocities How to set initial velocities depends on the physical case to be simulated Simplest case: thermal equilibrium: give initial Gaussian velocities to the atom (=Maxwell-Boltzmann distribution) Code snippet: ! Gaussian distributed velocities x(i) = vmean*gaussianrand(0) y(i) = vmean*gaussianrand(0) z(i) = vmean*gaussianrand(0) 1 3 Mean velocity obtained from 2 mv = 2 k T Basic physics note: if you start from perfect crystal 2 mean E pot = Ekin = E pot = 0 3 k BT 2 B positions, atoms initially have Epot = 0, whereas by the equipartition theorem, at a given temperature T, each 1 degree of freedom has 2 kBT of energy in both Epot and Ekin Hence in this case to get desired temperature T, you need to give initial velocities for 2T and simulate about 1 ps to get correct equipartitioning of Epot and Ekin Kai Nordlund, Department of Physics, University of Helsinki 13 MD method MD – Solving equations of motion The solution step r(i+1) = r(i) +v(i)Δt + 1/2 a Δt2 + correction terms is crucial What are the “correction steps”? There is any number of them, but the most used ones are of the predictor-corrector type way to solve differential equations numerically: Predictor stage: predict next atom positions: Move atoms: rp = r(i) + v(i) Δt + 1/2 a Δ t2 + more accurate terms Update velocities: vp = v(i) + a Δt + more accurate terms Get forces F = - ∇ V(rp) or F = F( Ψ(rp) ) and a = F/m Corrector stage: adjust atom positions based on new a: Move atoms: r(i+1) = rp + some function of (a, Δt) Update velocities: v(i+1) = vp + some function of (a, Δt) Kai Nordlund, Department of Physics, University of Helsinki 14 MD method MD – Solving equations of motion Simplest possible somewhat decent algorithm: velocity Verlet [L. Verlet, Phys. Rev. 159 (1967) 98] Another, much more accurate: Gear5, Martyna I recommend Gear5, Martyna-Tuckerman or other methods more accurate than Verlet – easier to check energy conservation [C. W. Gear, Numerical initial value problems in ordinary differential equations, Prentice-Hall 1971; Martyna and Tuckerman J. Chem Phys. 102 (1995) 8071] Kai Nordlund, Department of Physics, University of Helsinki 15 MD method MD – time step selection Time step selection is a crucial part of MD Choice of algorithm for solving equations of motion and time step are related Way too long time step: system completely unstable, “explodes” Too long time step: total energy in system not conserved Too short time step: waste of computer time Pretty good rule of thumb: the fastest-moving atom in a system should not be able to move more than 1/20 of the smallest interatomic distance per time step – about 0.1 Å typically Kai Nordlund, Department of Physics, University of Helsinki 16 MD method MD – Periodic boundary conditions A real lattice can be extremely big E.g. 1 cm of Cu: 2.1x1022 atoms => too much even for present-day computers Hence desirable to have MD cell as segment of bigger real system Standard solution: periodic boundary conditions This approach involves “copying” the simulation cell to each of the periodic directions (1–3) so that our initial system “sees” another system, exactly like itself, in each direction around it. So, we’ve created a virtual infinite crystal. Kai Nordlund, Department of Physics, University of Helsinki 17 MD method MD: periodics continued This has to also be accounted for in calculating distances for interactions “Minimum image condition”: select the nearest neighbour of an atom considering all possible 27 nearest cells Sounds tedious, but can in practice be implemented with a very simple comparison: Kai Nordlund, Department of Physics, University of Helsinki 18 MD method MD – Boundary conditions There are alternatives, though: Open boundaries = no boundary condition, atoms can flee freely to vacuum Obviously for surfaces Fixed boundaries: atoms fixed at boundary Unphysical, but sometimes needed for preventing a cell from moving or making sure pressure waves are not reflected over a periodic boundary Reflective boundaries: atoms reflected off boundary, “wall” Combinations of these for different purposes Kai Nordlund, Department of Physics, University of Helsinki 19 MD method MD – Temperature and pressure control Controlling temperature and pressure is often a crucial part of MD “Plain MD” without any T or P control is same as simulating NVE thermodynamic ensemble NVT ensemble simulation: temperature is controlled Many algorithms exist, Nosé, Berendsen, … Berendsen does not strictly speaking simulate thermodynamic NVT ensemble! – but is still often good enough NPT ensemble simulation: both temperature and pressure is controlled Many algorithms exist: Andersen, Nosé-Hoover, Berendsen Berendsen does not strictly speaking simulate thermodynamic NPT ensemble – but is still often good enough Kai Nordlund, Department of Physics, University of Helsinki 20 MD method Berendsen method equations In the Berendsen methods, the equations of motion are modified with a frictional term γ i mi vi as mai = Fi − γ i mi vi where the velocity scaling is shown [Berendsen, J. Chem. Phys. 81 (1984) 3685] to be achieved by a simple scaling of all velocities v → λv every time step with where T0 is the desired temperature and τT the temperature scaling time constant. In Berendsen pressure control, the system size and all atom coordinates are scaled with a similar factor Here β is the inverse of the bulk modulus of the system and τP the pressure scaling time constant. Kai Nordlund, Department of Physics, University of Helsinki 21 MD method Note on pressure control Never use pressure control if there is an open boundary in the system!! Why?? Think about it... Hint: surface tension and Young’s modulus Kai Nordlund, Department of Physics, University of Helsinki 22 MD method MD – cellular method and neighbour lists To speed up MD for large (> 100 or so) numbers of atoms, a combination of neighbour list and a cellular method to find the neighbours is usually crucial If one has N atoms, and want to find the neighbours for a finite-range potential, a direct search requires N2 operations – killing for large N Solution: if potential cutoff = rcut, divide atoms into boxes of size >= rcut, search for neighbours only among the neighbouring cells Neighbour list: form a list of neighbours within rcut+ rskin and update this only when needed Kai Nordlund, Department of Physics, University of Helsinki 23 MD method MD – cellular method and neighbour lists To speed up MD for large (> 100 or so) numbers of atoms, a combination of neighbour list and a cellular method to find the neighbours is usually crucial If one has N atoms, and want to find the neighbours for a finite-range potential, a direct search requires N2 operations – killing for large N Solution: if potential cutoff = rcut, divide atoms into boxes of size >= rcut, search for neighbours only among the neighbouring cells Neighbour list: form a list of neighbours within rcut+ rskin and update this only when needed Kai Nordlund, Department of Physics, University of Helsinki 24 Nonequilibrium extensions to MD Nonequilibrium extensions The basic MD algorithm is not suitable for many nonequilibrium simulations But over the last ~30 years augmentations of MD for nonequilibrium simulations have been developed Our group has specialized in irradiation effects and is deeply involved in this Here just a few aspects also relevant to other non- equilibrium systems described Kai Nordlund, Department of Physics, University of Helsinki 25 Nonequilibrium extensions to MD Variable time step schemes In case the velocities of atoms are varying during the simulation, it is worth using a variable time step Example for irradiation simulations Here Δxmax is the maximum allowed distance moved during any t (e.g. 0.1 Å), Δ Emax is the maximum allowed change in energy (e.g. 300 eV), vmax and Fmax are the highest speed and maximum force acting on any particle at t, respectively, cΔt prevents sudden large changes (e.g. 1.1), and tmax is the time step for the equilibrated system. This relatively simple algorithm has been demonstrated to be able to handle collisions with energies up to 1 GeV Kai Nordlund, Department of Physics, University of Helsinki [K. Nordlund, Comput. Mater. Sci. 3, 448 (1995)]. 26 Nonequilibrium extensions to MD Special boundary conditions In a zone which is far from thermodynamic equilibrium, one should not use any thermodynamic ensemble simulation method! Only use NVE = direct solution of Newton’s equation of motion, ”Free MD” Example for various irradiation cases: Kai Nordlund, Department of Physics, University of Helsinki 27 Interatomic potentials Whence the interactions? Recall from the MD algorithm: Get forces F = − ∇ V(r(i)) or F = F(Ψ) and a = F/m This is the crucial physics input of the algorithm! In the standard algorithm all else is numerical mathematics which can be handled in the standard cases to arbitrary accuracy with well-established methods (as outlined above) Forces can be obtained from many levels of theory: Quantum mechanical: Density-Functional Theory (DFT), Time-dependent Density Functional theory (TDDFT) - Limit: ~1000 atoms for DFT, ~100 atoms for TDDFT Classically: various interatomic potentials - Limit: ~ 100 million atoms! - Most relevant to irradiation effects Kai Nordlund, Department of Physics, University of Helsinki 28 Interatomic potentials Interatomic potentials In general, total energy of a system of N atoms can be written: Because the energy of a bond should not depend on the rij i j explicit atom coordinates, and if there are no external forces V1, this can be simplified to: θijk rik k From these the forces can be obtained in principle simply using In reality, doing and coding this derivative for already a 3-body potentials is incredibly tedious: it has to be perfect for energy conservation in the MD simulation Kai Nordlund, Department of Physics, University of Helsinki 29 Interatomic potentials Interatomic potentials Interatomic potentials can in general be divided in 2 classes: Molecular mechanics force fields Used in chemistry and biophysics and biology, with a few exceptions non-reactive, i.e. covalent chemical bonds cannot break - Also same element can have different interactions depending on place in molecule: Carbon-1, Carbon-2, … Useless in most materials science applications, since these typically involve phase changes, bond breaking etc. (Reactive) Interatomic potentials Chemical Only bonds can break and reform one atom type per element Kai Nordlund, Department of Physics, University of Helsinki 30 Interatomic potentials Reactive potentials For classical MD, the often only physical input is the potential Originally simple 2-body potentials, but by now these are almost completely out of use (except for noble gases and model studies) Dominant are 3-body potentials, and increasingly 4-body are used Two major classes of potentials: Tersoff-like: 1 Vi = Vrepulsive (rij ) + bijk (rij , rik ,θijk )Vattractive (rij ); bijk ∝ coordination of i neighbours Embedded-atom method-like (EAM) Vi = Vrepulsive (rij ) + Fi ρ (rij ) neighbours j Both can be motivated in the second momentum approximation of tight binding (“extended Hückel approximation” if you are a chemist) Related to Pauling’s theory of chemical binding [K. Albe, K. Nordlund, and R. S. Averback, Phys. Rev. B 65, 195124 (2002)] Kai Nordlund, Department of Physics, University of Helsinki 31 Interatomic potentials Embedded-atom method like potentials The general “EAM-like” functional form Vi = Vrepulsive (rij ) + Fi ρ (rij ) neighbours j is actually shared by a large number of different potentials A few descriptions of common forms: Finnis-Sinclair (1982) and Rosato group potentials: F is simply the square root function Motivated from 2nd moment tight binding Rosato group potentials: as above, but electron density just a simple exponential. Simple but often pretty good. Confusingly called “tight-binding potential” even though fully classical! Kai Nordlund, Department of Physics, University of Helsinki 32 Interatomic potentials Embedded-atom method like potentials The original EAM form by Foiles, Daw, Baskes ~1985: Vi = Vrepulsive (rij ) + Fi ρ (rij ) neighbours j Here F is the energy to embed an atom into an electron gas, and ρ an actual electron density function for neighbours j of atom i. Based on basic idea of metal as sea of free electrons with positive ion cores embedded in it EMT potentials: original potentials developed ~1980 from first principles, underlying the later EAM form. Developed mostly in Finland (Nieminen, Manninen, Puska) Kai Nordlund, Department of Physics, University of Helsinki 33 Interatomic potentials Embedded-atom method like potentials… And more potentials within same form Glue potentials (Ercolessi): pair part has also an attractive part CEM, Corrected Effective Medium theory potentials: completely different motivation, but can be squeezed to same form ‘Magnetic’ potentials (Dudarev, Derlet): recent addition of magnetic energy term into an EAM-like potential [Dudarev, Derlet, J. Phys.: Condens. Matter 17 (2005) 1] Finally, to enable description of covalent materials, Baskes et al extended the EAM form to include an angular term, “Modifed EAM” = MEAM potentials Kai Nordlund, Department of Physics, University of Helsinki 34 Interatomic potentials Interatomic potential development: general principles The general idea of fitting interatomic potentials is to adjust the parameters into data obtained experimentally or from deeper-level calculations (typically DFT) In case one gets as an end results a potential that can describe a wide range of properties outside the original fitting set, the potential can be considered transferable Example: Stillinger-Weber potential for Si: in spite of being fit in 1985 to only a small set of properties, it has turned out to describe well a wide range of other properties - Even very recently, this potential turned out to be the best in describing threshold displacement energies [Holmström et al, Phys. Rev. B 78, 045202 (2008)]. Set of properties selected for fit depends on field and aims Kai Nordlund, Department of Physics, University of Helsinki 35 Interatomic potential development Potentials developed in general In general, potentials suitable for irradiation effects exist: For almost all pure elements (mainly EAM-like for metals) For the stoichiometric state of a wide range of ionic materials - But these do not always treat the constituent elements sensibly, e.g. in many oxide potentials O-O interactions purely repulsive => predicts O2 cannot exist => segregation cannot be modelled For a big range of metal alloys Not so many potentials for mixed metal – covalent compounds, e.g. carbides, nitrides, oxides in non-ionic state Extremely few charge transfer potentials For organics only ReaxFF for CNOH, extended Brenner for COH systems NIST maintains a potential database, but pretty narrow coverage – one often really needs to dig deep in literature to find them Kai Nordlund, Department of Physics, University of Helsinki 36 Interatomic potential development Potential development aims (materials science point of view) First consider a potential for a pure element A. To be able to handle the effects described above, the potential should give: The correct ground state: cohesive energy, crystal structure etc. Describe all phases which may be relevant Describe melting well Describe defect energetics and structures well If we further consider irradiation of a compound AB: At high temperatures the compound may segregate, so we need good models for elements A and B separately! Fulfills all the requirements just given for a pure element Describes well the heat of mixing of the compound Describes defects involving atom types A and B well Kai Nordlund, Department of Physics, University of Helsinki 37 Interatomic potential development Potential development approach Achieving all this starts to sound prohibitively difficult But there is one common factor for the main requirements: Melting, defects and different phases all involve unusual atom coordination states Hence if we use a framework to fit as many coordination states of the system as possible, we have some hope of getting many of the properties right A Tersoff (Abell / Brenner)-like formalism can do this! This is our favourite formalism for mixed materials, and hence I will now describe it in detail Kai Nordlund, Department of Physics, University of Helsinki 38 Interatomic potential development Potential development approach We start by obtaining information on as many coordination states as possible: Usually at least: Z: 1 3 4 6 dimer graphite diamond SC 8 12 BCC FCC Data from experiments or DFT calculations Cohesive energy, lattice constant, bulk modulus for all Z Elastic constants for most important Fitting done in systematic approach introduced by Prof. Karsten Albe (TU Darmstadt) Kai Nordlund, Department of Physics, University of Helsinki 39 Interatomic potential development “Albe” fitting formalism Use Tersoff potential in Brenner form (unique mathematical transformation) The 3 parameters r0, D0 and β can be set directly from the experimental dimer interatomic distance, energy and vibration frequency! Kai Nordlund, Department of Physics, University of Helsinki 40 Interatomic potential development “Albe” fitting formalism In nn formulation, if material follows Pauling theory of chemical bonding, Log Energy/bond Key idea: DFT or expt. data dimer GRA DIA BCC SC FCC for all coordinations Bonding distance [Albe, Nordlund and Averback, Phys. Rev. B 65 (2002) 195124] Kai Nordlund, Department of Physics, University of Helsinki 41 Interatomic potential development “Albe” fitting formalism Pair-specific A-B interaction Three-body part modified from Tersoff form Second-moment approximation exponential No power of 3 ik-dependent angular term This form for bij conforms exactly to bijk ∝ 1 coordination of i [Albe, Nordlund and Averback, Phys. Rev. B 65 (2002) 195124] Kai Nordlund, Department of Physics, University of Helsinki 42 Interatomic potential development The “blood, sweat and tears” part There are all in all 11 parameters that must be specified Constructing a good potential means finding suitable values for these This is done by fitting to different experimental or densityfunctional theory values of ground state and hypothetical phases – also for other functional forms than Tersoff Not a trivial task! 1-2 years Kai Nordlund, Department of Physics, University of Helsinki [Schematic courtesy of Dr. Carolina Björkas] 43 Interatomic potential development Potentials developed by us We, and/or the Albe group, have so far developed potentials for: BN, PtC, GaAs, GaN, SiC, ZnO, FePt, BeWCH, FeCrC, FeCH, + He with pair potentials All these potentials include all the pure elements and combinations! Fitting code “pontifix” freely available, contact Paul Erhart Just to give a flavor of complexity that can be modelled: prolonged irradiation of WC by H and He Kai Nordlund, Department of Physics, University of Helsinki 44 Nanocluster simulations Specific example of MD uses: ways to simulate nanoclusters and their deposition A single nanocluster in vacuum is a prototypical nanoscience system … and not irrelevant experimentally: they can be made and studied in various high vacuum gas condensation setups’ Simulation of them is also a good case for MD: Limited number of atoms (a few tens to a few hundred thousand: easy with modern computers) -- fast In high vacuum, cluster does not interact with anything – just free NVE simulation describes it perfectly - Or to be precise, there is always electromagnetic (blackbody) radiation: but time scale for this is long, at least tens of microseconds at low temperatures. No boundary conditions needed Kai Nordlund, Department of Physics, University of Helsinki 45 Nanocluster simulations Nanocluster basics A nanocluster (nanoparticle): an agglomerate of some 10-100 000 atoms or molecules bonded together, where the local environment of each atom/molecule is similar Nanoparticle = any object in the 1 – 100 nm size range Nanocluster: a 1 – 100 nm agglomerate of identical atoms or molecules - A single molecule of complex structure is not a nanocluster Atom cluster ≈ nanocluster Nanocrystal = crystalline or polycrystalline nanocluster Several different ‘crystallographic’ shapes exist for clusters of same material Wulff polyhedron, single crystalline Kai Nordlund, Department of Physics, University of Helsinki Icosahedron, polycrystalline x 20 Decahedron, polycrystalline x 5 46 Nanocluster simulations How to make these nanocrystals in MD? Rule of thumb: CUT IT OUT!! Also true for nanowires & surface structures It is almost always much easier to cut out the shape from a bulk shape than to generate it from scratch Basic idea: - Loop over atom positions - Create atom i in FCC position - Check if it belongs inside crystal region - If it does, add it to array - End loop over atom positions Kai Nordlund, Department of Physics, University of Helsinki 47 Nanocluster simulations Checking of belonging Checking whether it belongs to the crystal is a nice small piece of vector algebra E.g. for Wulff polyhedron with 100 and 111 facets: - Set desired sizes in both directions: s100 and s111 based e.g. on energy optimization described above - Loop over atom positions - Create atom i in FCC position ri - Outside=0 - Loop over 100 facets: f = 100, 010, 001 - If (|ri· f| > s100 ) outside=1 - End loop over 100 facets - Loop over 111 facets: f = 111, -111, 1-11,11-1 - If (|ri· f| > s111 ) outside=1 - End loop over 111 facets - If outside==0, add it to array - End loop over atom positions Kai Nordlund, Department of Physics, University of Helsinki 48 Nanocluster simulations Actual code snippet Main loop from code which does this: Kai Nordlund, Department of Physics, University of Helsinki 49 Nanocluster simulations Making icosahedra & decahedra The icosahedral & multiply twinned decahedral clusters cannot be made this way, however They are NOT pieces cut out of FCC But one can use this approach to first make the constituent parts, then join them together after suitable rotation For instance multiply twinned icosahedron: - Make a perfect tetrahedron with side length corresponding to the desired size - Then make 20 copies, rotate them suitably around 20 = corner closest to origin and join together: Kai Nordlund, Department of Physics, University of Helsinki (111) 50 Nanocluster simulations Rotation Clusters coming from a cluster ion source or in a liquid solution will have a random rotation for sure! Hence after creation, they should be rotated randomly This is a simple piece of geometry, bearing in mind however how to choose a random angle in 3D with correct weighting of angles! Use e.g. Euler angles, form a rotation matrix and then rotate Kai Nordlund, Department of Physics, University of Helsinki 51 Nanocluster simulations Rotation – actual code for single vector Kai Nordlund, Department of Physics, University of Helsinki 52 Nanocluster simulations Thermalizing – or not? How to thermalize a nanocluster correctly is a nontrivial question Adding random velocities corresponding to desired temperature is trivial per se, as usual in MD However, what is the distribution of energy to translational, rotational and vibrational degrees of freedom?? This question does not have a unique answer!! … but answer will depend on how the nanocluster was synthesized Interesting side aspect: a nanocluster thermalized by random atom velocities will almost certainly be rotating If no rotations are desired, one should first determine axis of gyration, then the angular velocity, then subtract velocities corresponding to this one out of atoms But then energy is lost from system – is this correct? - Again - no unique answer! Kai Nordlund, Department of Physics, University of Helsinki 53 Nanocluster simulations Controlling nanocluster temperature and rotation After a cluster has been made, one might want to control its temperature and rotation But also here one has to be careful: At least the very common Berendsen T control has the artefact that it tends to speed up free nanocluster rotation - Obviously physically incorrect Nose-Hoover does not seem to have the same problem(?) Removing angular momentum is possible: - First one has to calculate the rotational axis, then subtract out the rotation - Can be made to work, but also cause artefacts: we recently got an error in a metal nanocluster melting point simulation of 100 K due to removing rotation On more than one occasion, the only solution that worked was setting the desired T initially (maybe over a few ps NVT simulation), then run in pure NVE ensemble with short enough time step to keep energy very well conserved Kai Nordlund, Department of Physics, University of Helsinki 54 Nanocluster simulations Direct simulation of nanocluster production An alternative approach to making clusters: simulate the formation directly! Fairly (or even fully) realistic Kinetic effects included Rotations, vibrations etc. will come automatically But quite time-consuming; long MD sim needed [E. Kesälä, A.Kuronen and K. Nordlund, Phys. Rev. B (2005) ] Kai Nordlund, Department of Physics, University of Helsinki 55 Nanocluster simulations Direct simulation of nanocluster production A more recent example: formation of mixed Si/Ag clusters Kai Nordlund, Department of Physics, University of Helsinki [J. Zhao, F. Djurabekova, K. Nordlund 2014] 56 Nanocluster simulations Direct simulation of nanocluster production Final shape may be all but spherical!! Example: Kai Nordlund, Department of Physics, University of Helsinki SiGe clusters with different Ge compositions 57 Nanocluster simulations Simulation of nanocluster deposition After the nc is created, simulating deposition/bombardment on a surface is easy Place it above and give desired kinetic energy towards it E=X meV/atom <hkl> Kai Nordlund, Department of Physics, University of Helsinki 58 Nanocluster simulations Example, with rotation and all Cu cluster on Cu at very low energy Kai Nordlund, Department of Physics, University of Helsinki [Meinander et al, Thin Solid Films 425, 297 (2002); Appl. Phys. Lett. 89, 253109 (2006)] 59 Nanocluster simulations: introduction to tutorial Introduction to tutorial simulations A classic question in physics is how the melting temperature of nanoclusters depends on their size Theory predicts a smooth behaviour with cluster melting point going down with diameter due to weakening from surface [original derivation: Area Volume C r2 − C2 3 = Tmelt ,bulk − 2 r r Tmelt (r ) = Tmelt ,bulk − C1 P. Pawlow Z. Phys. Chem. 1909]: (C1, C2 are constants determined from surface/interface energy) = Tmelt ,bulk Experimental confirmation came in 1976 [Buffat, Borell, Phys. Rev. A 13, 2287] Kai Nordlund, Department of Physics, University of Helsinki 60 Nanocluster simulations: introduction to tutorial How to set up a nanocluster melting simulation? Remember the note on possible T control artefacts above! Safest approach: Create Set nanocluster of desired size (e.g. d=5 nm) the initial temperature of the atoms using the Maxwell- Boltzmann initialization (see above) for T = 2xdesired T Run simulation of nanocluster in pure NVE ensemble (no T or P control) in cell with periodic boundary conditions much larger than cell diameter Determine actual temperature of cell by averaging over a large number of time steps Monitor visually or some analysis method whether it melts Repeat using binary search in a wide T range until you find the melting point with desired accuracy Compare Kai Nordlund, Department of Physics, University of Helsinki with experiments and literature 61 Nanocluster simulations: introduction to tutorial Recent example Test of melting point of a new bond-order potential for Au [M. Backman, N. Juslin, and K. Nordlund, Bond order potential for gold, Eur. Phys. J. B 85, 317 (2012)] Kai Nordlund, Department of Physics, University of Helsinki 62 Molecular statics Molecular statics: definition In some applications one wants just to find the ground state atom coordinates of a system, i.e. minimize the total energy to the nearest energy minimum. This is called molecular statics Note: now we do not discuss trying to find the global energy minimum, which is still another task If you have an MD code, you can always do this by simply scaling the T in the simulation slowly down to 0, and run MD However, this is not particularly efficient approach, since the atoms vibrate due to the T (velocities) There are more efficient ways for this, in particular conjugate gradients (also widely used in DFT) Kai Nordlund, Department of Physics, University of Helsinki 63 Molecular statics Steepest descent energy minimization Before we describe conjugate gradients, let’s look at the obvious approach: steepest decent energy minimization Simply always move down the energy gradient = force Obviously will lead to local energy minimization Algorithm: Kai Nordlund, Department of Physics, University of Helsinki 64 Molecular statics Line minimization Step 3 in the algorithm above is a so called “line minimization”: simply move in the given vector (force) direction until you find a minimum Standard approach: “minimum bracketing”, Brents method If interested in details, see Press et al, “Numerical recipes” Kai Nordlund, Department of Physics, University of Helsinki 65 Molecular statics Conjugate gradient energy minimization Steepest decent has a problem, however: it tends to lead to a zig-zag motion, inefficient for energy minimization The “Conjugate gradient” approach is designed to overcome this problem Basic idea: use instead of only force, two vectors g and h constructed in a way to avoid the zig-zag motion, by being “conjugate” to each other. Can be achieved fairly simply: Kai Nordlund, Department of Physics, University of Helsinki 66 Molecular statics Conjugate gradients (CG) algorithm In algorithmic form, this becomes only slightly more complicated than steepest descent: If interested in details, see Press et al, “Numerical recipes” Kai Nordlund, Department of Physics, University of Helsinki 67 Molecular statics Adaptive conjugate gradients (ACG) – optimized for interatomic potentials… However, in MD the force/energy calculation is the most time-consuming part, and hence the line minimization is fairly heavy. Nordlund et al developed a scheme that avoids the line bracketing by using the knowledge that for an atomic system, “the minimum is out there”. Step size selected with an educated guess of typical shapes of interatomic potentials Increase If step size slowly energy goes up, return to previous position and reduce step size x 2 Speedup of 3-4 compared to normal CG Kai Nordlund, Department of Physics, University of Helsinki 68 Molecular statics Adaptive conjugate gradients (ACG) algorithm [Nordlund et al, J. Appl. Phys. 88, 2278 (2000)] Kai Nordlund, Department of Physics, University of Helsinki 69 Molecular statics Example: strain field for Co nanocluster in Cu Example of results: analysis of strain field for Co nanoclusters in Cu Note: displacements from perfect positions exaggerated by a factor of 3 [C. Zimmermann, M. Yeadon, M. Ghaly, K. Nordlund, J. M. Gibson, R. S. Averback, U. Herr, and K. Samwer, Phys. Rev. Lett. 83, 1163 (1999)] Kai Nordlund, Department of Physics, University of Helsinki 70 Molecular statics MD relaxation The velocities available in MD can actually be used to achieve fairly fast relaxation by a simple trick: Start at any temperature T you may be. The switch away from any other temperature control, and instead for every atom, form the dot product of the velocity and force: d = Fi ⋅ vi If the dot product is negative, atom is moving up the force. In this case, set vi = 0 This way all atoms are guaranteed to move downwards in force This approach may be good to get close to minimum fast, but for very accurate minimization, CG or ACG is likely still more efficient Kai Nordlund, Department of Physics, University of Helsinki 71 The end: practical advices “Big” MD codes MD codes in wide use: LAMMPS = Sandia national labs parallel MD code. By far the most used materials science code. Open source, free, well documented, big user base. Works like a script-like language. Only possible downside “creeping featurism” - Used now in tutorial DL_POLY = Popular fairly widely used code from the UK. Works for both inorganic and organic materials. Gromacs = Dominating MD code for biophysics simulations. Only supports molecular mechanics force fields, so not suitable for materials science More info on any of these: just google on the name Kai Nordlund, Department of Physics, University of Helsinki 72 The end: practical advices Further reading General: Classic book: Allen-Tildesley, Molecular dynamics simulations of liquids, Oxford University Press, 1989. Newer book: Daan Frenkel and Berend Smit. Understanding molecular simulation: from algoritms to applications. Academic Press, San Diego, second edition, 2002 Ion irradiation-specific reviews: K. Nordlund and F. Djurabekova, Multiscale modelling of irradiation in nanostructures, J. Comput. Electr. 13, 122 (2014). K. Nordlund, C. Björkas, T. Ahlgren, A. Lasa, and A. E. Sand, Multiscale modelling of Irradiation effects in fusion reactor conditions, J. Phys. D: Appl. Phys. 47, 224018 (2014). Kai Nordlund, Department of Physics, University of Helsinki 73 Examples of MD modelling results Research examples Time permitting… Kai Nordlund, Department of Physics, University of Helsinki 74 Examples of MD modelling results 1. Chemical sputtering of metals Using classical molecular dynamics and experiments, we showed that pure metals can sputter chemically − Light metals under H bombardment sputter by breaking bonds! [Björkas et al, New J. Phys. 11, 123017 (2009); Nordlund et al, NIM B 269, 1257 (2011)] 75 Kai Nordlund, Department of Physics, University of Helsinki 75 Examples of MD modelling results 2. Swift heavy ion effects on materials Swift heavy ions (Ekin > 100 keV/amu) can be used to fabricate and modify nanostructures in materials We are using multiscale modelling of electronic energy transfer into atom dynamics to determine the effects of swift heavy ions on materials We have explained the mechanism by which swift heavy ions create nanostructures in silicon, silica and germanium and change nanocrystal shapes [Ridgway et al, Phys. Rev. Lett. 110, 245502 (2013); Leino et al, Mater. Res. Lett. (2013)] Kai Nordlund, Department of Physics, University of Helsinki 76 Examples of MD modelling results 3. Surface nanostructuring Together with Harvard University, we are examining the fundamental mechanisms of why prolonged ion irradiation of surfaces leads to formation of ripples (wave-like nanostructures) We overturned the old paradigm that ripples are produced by sputtering and showed ab initio that they can in fact be produced by atom displacements in the sample alone Kai Nordlund, Department of Physics, University of Helsinki [Norris et al, Nature commun. 2 (2011) 276] 77 Examples of MD modelling results 4. Modelling of arc cratering We have developed new concurrent multiscale modelling methods for treating very high electric fields at surfaces Using it we are examining with a comprehensive multiscale model the onset of vacuum electric breakdown We have shown that the complex crater shapes observed in experiments can be explained as a plasma ion irradiation effect – multiple overlapping heat spikes [H. Timko et al, Phys. Rev. B 81 (2010), 184109] Kai Nordlund, Department of Physics, University of Helsinki 78 Examples of MD modelling results 5. Cluster cratering over 40 orders of magnitude Using classical MD, we demonstrated that at a size ~ 10000 atoms, cluster bombardment starts producing craters with the same mechanism as meteorites on planets − 100 million atom simulations with statistics [J. Samela and K. Nordlund, Phys. Rev. Lett. 101 (2008) 27601, and cover of issue 2] Kai Nordlund, Department of Physics, University of Helsinki 79 Summary Molecular dynamics is a powerful tool – when you know what you are doing! And a lot of fun! Kai Nordlund, Department of Physics, University of Helsinki 80 Kai Nordlund, Department of Physics, University of Helsinki 81