Surfaces

Definitions: An interface is the area that is between two phases. It should be noted that the phases do not need to be in different states of aggregation. Geometrically, a surface is the collectivity of all areas confining an object to the outside. Chemically, a surface is a two-dimensional alignment of atoms where the chemical potential is different to the chemical potential of a correspondent (three-dimensional) bulk structure. Surfaces are always interfaces. In computational chemistry, surfaces are often defined as the atomic layer(s) between the bulk and a cavity.

Ideal surfaces are obtained by cutting a plane through an ideal crystal. They often differ from real surfaces as Atoms are not aligned ideally Crystal growth is not ideal in reality Atoms at the surface are not co-ordinated completely The lattice may rearrange to account for the above effects The environment interacts with the surface

Surface energy is the change of internal energy needed to create a surface: ( ) Surface free energy is the same, but from an free energy perspective where free energy is defined as (with constant volume and temperature)

Surface free energy is thus ( ) Internal energy is the change of energy when the internal composition of a system is changed. Free energy is the energy needed to generate a system in thermal equilibrium with its environment. With some applied thermodynamics we can show that under constant volume and temperature enthalpies and energies are equal, thus and

Surface energies can be computed. Within a constant volume simulation cell (bulks + cavities) energies are computed with different surface areas (cf. figure)

The surface energies can then be computed using a simple algorithm: I. II. III. IV. create both structures calculate E normalise E to the number of atoms (N) in the surface: E = E/N normalise E to the surface created (per surface atom): s = E/A

Sorption is a sum of reversible physical and chemical processes an accumulation of a substance in a phase (absorption) or interface (adsorption) adsorption by chemical binding (chemisorption) adsorption by physical binding (physisorption)

Materials in the Virtual Lab

The virtual lab is a computational modelling environment (this means that no lab experiments are done) uses model potentials for atomic interactions Ab initio methods Provide a functional relation between atomic positions, electronic structure and potential energy of the system Neural networks Provide a functional relation between atomic positions and potential energy of the system

Force fields Provide a functional relation between atomic positions and potential energy of the system Parameters derived empirically

For chemical problems feedEnergy depends solely on forward neural networks are Ideal potentials are used electronic structure (which is a used (harmonic oscillator, Lennard- functional of atomic positions) Jones, Morse oscillator, cubic Several layers containing and quartic terms for oscillators Explicit treatment of electronic artificial neurons (nodes) etc.) structure and related properties Output layer: Single node Parameters only for recognized containing the energy potentials Approximations used (HartreeFock Theory, Density Functional Hidden layers between input Theory, Perturbation Theory, and output layer have no Coupled Cluster Theory etc.) physical meaning, but define the functional form oft he neural network. The more hidden layers, the more flexible is the neural network Scale N2 to N4 depending on DFT scales n2.5 to n4 depending forcefield class on functional and method, highly correlated methods like coupled cluster methods scale up to n8 or higher Neural networks: Training set calculations scale like the method used times number of calculations needed to teach the network, once parametrized it scales similar to force fields Accuracy of at least 0.1 kJ/mol Accuracy similar to that of (more precise: 1 meV) in training set calculations standard applications, significantly better if highly correlated methods are used Treatment of a few thousand Treatment of a few million atoms possible, short atoms possible, long trajectrajectories (up to nano- tories (up to seconds), accurate seconds), accurate description description of bond breaking of bond breaking and bond and bond formation and no formation, no structural pre- structural prejudice by paramejudice by parametrization trization (if both are included in training set) 3

Can reach accuracy of a few kcal/mol

Treatment of a few million atoms possible, long trajectories (up to seconds), cannot describe structures unforeseen in parametrization, difficulties with breaking and formation of chemical bonds

A general feed-forward (information flows only in one direction from input layer to output layer) neural network is presented on the left. A corresponding neural network for chemical applications is presented on the right. The Cartesian coordinates of the atoms are used as input. The neural network transforms the overall positions into bond length, angles etc. using symmetry functions. This allows to form clusters (atomic neural networks) for the computation of atomic energies of the central atoms of each cluster. All atoms are central atoms for one cluster each. Finally, the atomic energies are summed up and give the total energy. The atomic energy of a central atom of a cluster is trained before by setting up environments within van der Waals-radius of the respective atom. Each different environment represents a different, but possible, arrangement of atoms around the central atom, e.g. the central atom in the bulk in ideal surrounding, the central atom in the bulk next to a defect site, the central atom on the surface etc. The more complete the set of environment is in the training set, the more precise and flexible is the neural network. However, adaptive neural networks are also possible where the neural network requests a new environment calculation if it runs into an unknown environment.