Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. ATOMISTIC SIMULATION METHODS FOR STUDYING SELF HEALING MECHANISMS IN Al/Al2O3 Ivan Lazic, Marius Ernst and Barend Thijsse Department of Materials Science and Engineering, Delft university of Technology +31 (015) 2789518 +31 (015) 2786730 I.Lazic@tudelft.nl www.structureandchange.3me.tudelft.nl New methods for modeling initial oxide film growth using MD (Molecular Dynamics) simulations are explored, in order to study the atomic mechanism of self-healing oxidation in metal/oxide systems, in particular in Al/Al2O3. Computationally, this is a challenging undertaking. A new version of the MEAM (Modified Embedded Atom Method) potential is proposed and extended by a variable charge ionic potential model from the literature. This makes it possible to include angular forces, which are needed for relatively open crystal structures such as metal oxides, as well as localized and dynamic charge transfer between Al and O, which is necessary to handle time-dependent local composition variations. As Coulomb solver we use the PPPM method (particle-particle-particle-mesh). In this work the first results are reported. Tests of the ionic potential model in combination with the PPPM solver yield excellent results. Simulations of oxygen atom arrivals at an Al surface will be presented. Keywords: molecular dynamics, oxides, self-healing 1 Introduction The surface oxides on aluminum and aluminum alloys can be called “self-healing” in that they quickly re-form after scratching. The initial stage of oxidation is very rapid, much faster than for a more noble metal such as e.g. Ru (Fig. 1) and is therefore difficult to follow experimentally. Here simulations would be able to extend time scales and resolution. Such self-healing phenomena of metal oxides using Molecular Dynamics (MD) simulations require sophisticated interatomic potentials of the underlying metal-oxygen systems, which is a challenging undertaking. The main target of our investigation is Al2O3. For oxides and metal/metal-oxide systems in general, extra difficulties appear because of the presence of ionic bonds, giving rise to long-range Coulomb interactions. Moreover, at interfaces, surfaces and defects, the atomic charges cannot be considered fixed, and models that allow charge transfer should be applied. Finally, aluminum oxide has a relatively open structure, which suggests that angular terms in the interatomic potentials could play an important role. For these reasons we are developing a potential for Al/Al2O3 that combines a charge-transfer potential for the electrostatic interactions with the Modified Embedded Atom Method (MEAM) [1] for the non-electrostatic interactions. Such a potential would also be applicable to the relatively open oxides of e.g. Si and to the mildly ionic III-V and II-VI semiconductors. This paper reports on the potential under development and on the implementation of the Coulomb solver. 1 © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. Previous MD work on Al/Al2O3 has been done using MEAM but without considering charges explicitly [2], and using charge transfer ionic potentials but in combination with EAM, which does not includes angular dependencies [3][4]. Here we propose to use the Charge Transfer Ionic Potential (CTIP) published in ref. [4] in combination with novel version of the MEAM. Unlike previous work, we use the iterative Particle-Particle-Particle-Mesh (PPPM) method as Coulomb solver [5]. PPPM easily adapts to periodic and nonperiodic boundary conditions, and, because of its localized computability, it is convenient for code parallelization. An additional reason is that the computational efficiency of PPPM is ϑ ( N log N ) while, e.g., for Ewald Summation it is ϑ ( N 3 / 2 ) [6]. A disadvantage of PPPM is that the maximal error can not be predicted analytically. Figure 1: Oxide growth measured by in-situ ellipsometry on oxidizing single crystals of aluminum and ruthenium. Formation of the initial, chemisorbed oxide layer is easily noticeable for ruthenium but not for aluminum. Very likely it does take place also in aluminum but very fast 2 Theory and methods In our MEAM+CTIP model, the energy U of a system of N atoms is the sum of a nonelectrostatic and an electrostatic part, U = U nes +U es , where the non-electrostatic MEAM part is given by N N N U nes = ∑ ∑ 12Φ IJ (rij ) + ∑ FI (xi ) (1) i=1 i=1 j=1 j≠i in which upper case I, J denote the chemical types of the atoms i, j. The pair potential and embedding functions for each of the chemical types or type combinations are given by Φ(r) = −E p (1+ η + c pη 2 + d pη 2 )e−η , 2 η = α p (r / rp −1), (2) © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. F(x) = Ae x ln x + Be x + Ce x , (3) where the p and e subsripts denote adjustable parameters. The quantity xi is proportional to the square of what in (M)EAM terminology is called the “background electron density at atom i” and contains all dependencies on the i-angle in the ijk atom triplets, 3 xi = ∑ t I(l ) ∑ ∑ p(lJ )e l=0 −qJ(l ) rij pK(l )e−qK (l ) rik P (l ) (cosθ jik ) . (4) j≠i k≠i Here, t(l), p(l), q(l) are fitparameters for each chemical type, and P(l) is the Legendre polynomial. For simplicity, Eqs. (1-4) are given without angular and radial cutoff functions. New in this MEAM format is that the embedding function has been extended by two terms, that the concept of “reference structure” has been abandoned, and that the pair potential is a parametrized (Morse-like) function rather than a function determined by a prescribed equation of state of a particular crystal phase. Eqs. (1-4) are more flexible than the classical format, without changing much in the underlying physical picture. With these expressions it proved possible to represent the well-known Stillinger-Weber and Tersoff-III silicon potentials to a very high degree of accuracy [7], which suggests that they will be appropriate to the current Al/Al2O3 case as well. In our implementation of the electrostatic CTIP part, the long-range Coulomb interactions are handled by the PPPM method which is briefly explained in Figs. 2 and 3. The grids on the figures represent the mesh, which holds the charge distribution in the middle panel of Fig. 3. Assuming pointlike charges, the interaction between a charged atom i and all other charged atom j separated by rij, including all periodic images, is divided into two parts by adding and subtracting a charge distribution ρ j (r) around every atom j. The volume integral of this distribution must be equal to qj. If it is chosen to be Gaussian, ρ j (r) = q jσ −3π −3/2 exp(−r 2 / σ 2 ) , it can be proven that the interaction energy between point charge i and point-charged atom j minus its own charge distribution is qi q j erfc(rij / σ ) / rij . The complementary error function decreases rapidly with distance, which means that this part of the interaction has become short-ranged and a cutoff radius can be used (Fig. 2), normally on the order of 5 Å. The rest of the interaction (Fig. 3) is calculated by placing point charge i in the electrostatic potential on the mesh, solved by the Poisson equation ∇ 2ϕ mesh = −ρ mesh / ε 0 . Here, ρ mesh is the total charge of all individual charge distributions ρ j (r ) , plus i’s own charge distribution, allocated to the mesh points through a diffusion-like algorithm. In this way ρ mesh is the same for all atoms i and the Poisson equation needs to be solved only once. For each atom, the spurious interaction between its point charge and its charge distribution can then be subtracted analytically (Fig. 2, right). In summary, one obtains N qi q j j =1 j ≠i rij ∑ kc N qi q j j =1 j ≠i rij = ∑ kc ⎛ rij erfc⎜⎜ ⎝σ ⎞ ⎟ + q i (ϕ mesh ( q ) − ϕ self ( q i )) ⎟ ⎠ 3 (5) © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. Figure 2: Short-range part (“PP”) of the interaction between atom i and all other atoms, which is taken into account directly, showing the cutoff radius. The mesh is not used here Figure 3: The rest of the interaction (“PM”). The middle frame shows the interaction of i with the charges attributed to the mesh points. The Poisson equation is used here. The frame on the right shows the correction for self-interaction where kc = 1 / 4 πε 0 , and ϕ self (qi ) = 2kc qi / σ π is the self-electrostatic potential. The vector notation indicates the set of all charges. The mesh size should be no more than one quarter of the cutoff radius. According to the CTIP model the charges in the system are not constant. Every atom is allowed to become ionized by charge transfer to and from other atoms. The ionization energy Eiionization (qi ) = χ i qi + J i qi2 / 2 , with χ i the electronegativity and Ji the electrostatic hardness of atom i, is compensated by the appearance of Coulomb electrostatic energy 1 EiCoulomb = k c ∑ qi q j / rij . The charges that atoms obtain for a certain atomic arrangement are 2 determined by searching for the minimum of total electrostatic energy Ues (ionization plus Coulomb) with the condition of total system charge neutrality. This energy is given by N 1 ⎛ ⎞ 1 N N U es = ∑ ⎜ χ i q i + J i q i2 ⎟ + ∑ ∑ (k c q i Z j ([ j | f i ] − [ f i | f j ]) + k c q j Z i ([i | f j ] − [ f i | f j ]) ) + 2 ⎠ 2 i =1 j =1 i =1 ⎝ j ≠i ⎛⎛ N q i − q min,i 1 N N k c q i q j [ f i | f j ] + ∑ ω ⎜ ⎜1 − ∑ ∑ ⎜⎜ 2 i =1 j =1 q i − q min,i i =1 ⎝⎝ j ≠i ⎞ ⎛ ⎟(q − q ) 2 + ⎜1 − q max,i − q i i min,i ⎟ ⎜ q i − q max,i ⎠ ⎝ 4 (6) ⎞ ⎞ 2 ⎟ ⎟( q − q i max,i ) ⎟ ⎟ ⎠ ⎠ © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. where notations [a | f b ] and [ f a | f b ] (a=i,j; b=i,j; a≠b) denote the Coulomb interaction f (r , q ) f (r , q ) f (r , q ) integrals [a | fb ] = kc ∫ b b b dVb and [ f a | f b ] = k c ∫ ∫ a a a b b b dVa dVb . This is rvv rav V V V a b b according to the used charge model, in which each atom has a pointlike charge part and a distributed part: ρ iatom (r ) = Z i δ (r ) + (qi − Z i ) f i (r ) . For mathematical simplicity the distributed part of the atomic charge is chosen to be f i ( r ) = ξ i3 exp( −2ξ i r ) / π . The last term in Eq. (6) is an extra term that softly bounds the charges between the values q min and q max ( ω > 0 ) to account for chemical valence. Parameters that need to be specified for each type of atom in the system are then: χ i , J i , Z i , ξ i , q min , and qmax . After applying PPPM (Eq. (1)) to Eq. (6) we achieve U es = ⎛⎛ q − qmin,i 1 N 2ω ⎜ ⎜1 − i ∑ 2 i =1 ⎜ ⎜⎝ qi − qmin,i ⎝ ⎞ ⎛ ⎟(q − q ) 2 + ⎜1 − qmax,i − qi min, i ⎟ i ⎜ qi − qmax,i ⎝ ⎠ ( ) N ⎞ ⎞ ⎟(q − q ) 2 ⎟ + 1 ∑ 2 χ q + J q 2 + q ⋅ (ϕ (q ) − ϕ (q )) + i i i i i i mesh self i max, i ⎟ ⎟ 2 i =1 ⎠ ⎠ N N ⎛ ⎞ 1 N N (kc qi Z j ([ j | fi ] − [ fi | f j ]) + kc q j Zi ([i | f j ] − [ fi | f j ])) + 1 ∑∑ kc qi q j ⎜⎜ 1 erfc⎛⎜⎜ rij ⎞⎟⎟ + ([ fi | f j ] − 1 ) ⎟⎟ ∑∑ 2 i =1 j =1 2 i =1 j =1 rij ⎠ ⎝σ ⎠ ⎝ rij j ≠i j ≠i (7) The minimization of Ues together with the condition ∑ N i =1 q i = 0 is done using the Conjugate- Gradient method [8]. For the line minimization in the charge space the Brent algorithm was used [8]. In order to test the accuracy of the CTIP/PPPM implementation, we use the fact that the electrostatic energy for the simple case of, e.g., a NaCl crystal type can also be determined and minimized analytically, in terms of the Madelung constant, N 1 ⎛ ⎞ 1 N N U es (q ) = ∑ ⎜ χ i qi + J i qi2 ⎟ + ∑∑ (kc qi Z j ([ j | f i ] − [ f i | f j ]) + kc q j Z i ([i | f j ] − [ f i | f j ]) ) + 2 ⎠ 2 i =1 j =1 i =1 ⎝ j ≠i N N (8) 2 1 1 N NaCl q kc qi q j ([ f i | f j ] − ) − α madelung ∑∑ 2 i =1 j =1 rij 2 r0 j ≠i NaCl where qi = q if atom i is Na and q i = − q if atom i is Cl, α madelung is the Madelung constant for the NaCl crystal type, and r0 nearest neighbor distance. The double summation terms include only short-range functions and converge very fast. In this formula the charge bounding energy term is not included for simplicity. 3 Results (1) Results of how well PPPM works are given in Table I. The basic tests of Madelung constant calculations were done for several artificially made but convenient structures. 5 © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. If a unit point charge is placed inside the heavily drawn box (at fractional coordinates X, Y, Z) in Fig. 4 and a mirroring operation is applied, infinite three-dimensional structures as shown in Fig. 4 can be generated. If the charge is placed in the middle of the box we achieve exactly the NaCl crystal type, but by varying the coordinates of its position inside the box we obtain other types. For all of these we can calculate the Madelung constant exactly. A comparison of the exact and the PPPM results for the Madelung constant is shown in Table I. Figure 4: Generating the crystal structure samples for the Madelung constant test Table I: Madelung constants for different structures calculated exactly and using PPPM X 0.2 0.6 0.4 0.1 0.5 0.8 0.7 0.9 0.6 Y 0.5 0.6 0.9 0.1 0.5 0.2 0.9 0.7 0.5 Z 0.3 0.1 0.8 0.7 0.5 0.5 0.8 0.1 0.4 Madelung constant Exact PPPM 2.83447 2.83447 5.10754 5.10755 5.27626 5.27628 6.47665 6.47664 1.74755 1.74756 3.26069 3.26071 5.29997 5.29998 6.47665 6.47664 1.85929 1.85934 (2) Values of charges and potential energies per atom for the NaCl structure in which one allows charge transfer can be calculated exactly (Eq. 8) and compared with the minimization carried out by PPPM (Eq. (7)). Results are given in Fig. 5. The curves “exact” and “PPPM ( ω = 0 )” overlap almost perfectly. As expected, the PPPM calculations with bound charges (in our case ω = 20 eV/e2 ) yield different curves in the region where charge bounding is actually taking place. The parameters used for these calculations are shown in Table II. These parameters are just test parameters, not literature values or values achieved by fitting. cut −off The cut-off radius for all short-range terms in PPPM method was chosen to be rpppm = 5.4 Å. 6 © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. Figure 5: Charges q (left) and electrostatic potential energy per atom Ues/N (right) for the NaCl crystal structure type after energy minimization calculated analytically and using PPPM for different nearest neighbor distances. The inserts show the deviation between PPPM (ω = 0) and the exact method Table II.: Test parameters for Na and Cl atomic types Element q q max χ[eV] J [eV ] ξ [Å-1] Z min Na Cl 0 -1 1 0 -3.4 2.0 10 14 1.0 2.1 0.6 0.0 (3) The effect of angular forces is illustrated by a simplified calculation of the III-V compound AlP. Using a realistic MEAM potential, the cubic ZnS phase is the stable crystal phase. By changing only one parameter value in the MEAM potential, namely t(3) in Eq. 4 , which controls the energy associated with the third spherical harmonic describing an atom’s geometrical environment, the cubic CuAu phase can be made the stable crystal phase (Fig. 6). Figure 6: Energies of two phases of a simplified model of AlP as a function of nearest-neighbor distance. By simply changing a single parameter in the MEAM potential, the relative stability of the two phases can be tuned (left and right). Charge transfer varies from approximately 0.25 e at small R to 0.13 e at large R 7 © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. (4) The first simulation results of an oxygen atom approaching an Al surface kept at 375 K are shown in Fig. 7. The oxygen atom was directed at the Al sample at normal incidence, with an initial kinetic energy of 0.0125 eV, which corresponds to the average thermal speed at 300 K. For this simulation, the EAM potential of ref. 4 were used, i.e. a potential without the angular terms l = 1, 2, 3 in Eq. (4). After impact, the oxygen atom reaches the second Al atomic monolayer from the surface and stays there as an interstitial. As expected, no permanent change was produced in the Al lattice structure. The amount of charge that the oxygen atom has acquired as interstitial is –0.54 e, which is far away from the minimum value of –2 e, and this charge is almost totally transferred to the six closest neighboring Al atoms (around +0.08 e each). Figure 7: Oxygen atom approaching Al surface held at 375 K. a) Oxygen atom reaching the first Al plane. b) Final oxygen position as interstitial. Upper two figures: view along the Al <110> direction. Lower two figures: top view along the Al <001> direction. The squares indicate the same Al unit cell. Colors represent different charge ranges: light blue (–0.02 e, +0.02 e), dark blue (+0.07 e, 0.25 e), magenta (–0.55 e, –0.45 e) 4 Discussion Results of calculating Madelung constant using PPPM, given in Table I, are in extremely good agreement with analytically determined values for all structures generated as described above (Fig. 4). In agreement with this, the results of using PPPM as a long-range Coulomb solver in minimization process are also very satisfactory (Fig. 5); the deviations using PPPM in comparison with the exact calculations do not exceed 8 meV/atom for the electrostatic potential energy and are less than 0.009 e for the charges in the NaCl crystal type. 8 © Springer 2007 Proceedings of the First International Conference on Self Healing Materials 18-20 April 2007, Noordwijk aan Zee, The Netherlands Ivan lazic et al. In the range of nearest neighbor distances where the equilibrium positions are expected, between 2 and 3.5 Å, these errors are even less then 1 meV/atom and 0.001 e. Using an MEAM instead of EAM potential, the possibility of fine tuning of nonelectrostatic angular dependent part of the potential interaction should allow us better fitting to experimental and ab initio data for open structures as Al2O3. Finally, the approach of a single oxygen atom towards the surface of a pure Al crystal gives us some insight in very beginning of process of aluminum oxidation. Further work is needed to understand why the interstitial oxygen atom in Fig. 7 is only partially charged, while one could expect a full charge of –2 e. This may be due to a flaw in the charge transfer model, although it correctly leads to fully charged atoms in Al2O3. 5 Conclusions In this work the first steps towards building a sophisticated enough simulation system for metal-oxide/metal-alloy systems and some first results are reported. The combination of a charge transfer ionic potential (CTIP) [4] together with PPPM as long-range Coulomb interaction solver was tested against analytical calculations for simple system such as the NaCl crystal type. The results for charges and energies are excellent. Also, a modified version of the MEAM potential is proposed, one that is designed for more flexibility then the original one. Control of the phase stability by adjusting parameters of the angular dependent terms in the model has been demonstrated to work well, even together with atomic charges and charge transfer in the system. A first simulation on oxygen atom approaching an Al surface yields realistic results on initial oxidation. These first tests will be followed by further simulations that combine MEAM and CTIP in the studying of self-healing effect in Al/Al2O3. ACKNOWLEDGMENTS The authors would like to thank Prof. Dr. Marcel Sluiter, Darko Simonovic, MEng, and Dr. Wim Sloof from the Department of Material Science and Engineering, Delft University of Technology, for discussions and critical review of the manuscript. This research is supported by the Netherlands Foundation for Fundamental Research on Matter (FOM) and the Netherlands Institute for Metal Research (NIMR) within the project 02EMM31. REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. M. I. Baskes, Phys. Rev. B46, 2727 (1992). M. I. Baskes, Modified Embedded Atom Method Calculations of interfaces, Sandia National Laboratories Livermore, CA 94551-0969, Edited by S. Nishijuna and H. Ondera, (1996). T. Campbell, R.K. Kalia, A. Nakano, P. Vashishta, S. Ogata, and S. Rodgers, Phys. Rev. Lett. 82, 4866 (1999). X. W. Zhou, H. N. G. Wadley, J. S. Filhol, M.N. Neurock, Phys. Rev. B69, 035402 (2004). R. W. Hockney, J. W. Eastwood, Computer Simulation Using Particles, (McGraw-Hill, New York, 1981). R. Fletcher and C. M. Reeves, Computer J. (UK) 7, 149 (1964). M. Timonova, I. Lazic, and B.J. Thijsse, to be published. W. H. Press, S. A. Teukolsky, W. T. Vetterling and B. P. Flannery, Numerical Recipes, Chapter 10, Cambridge University Press, 2nd edition, (1992 9 © Springer 2007