Adrien Poissier

PhD student

Address :
46 allee d'Italie
69364 LYON

Affiliation :
Centre Europeen de Calcul Atomique et Moleculaire

Contact :
Tel : +33 4 7272 8640
Fax : +33 4 7272 8636

I. RESEARCH HIGHLIGHT

The interaction between water and metal surfaces are on first importance in the material science world. At the nano scale the physical properties are transformed and governed by the quantum effects. The quantum theory is computationally heavy and requires a lot of time and memory in computers even for relatively small calculations. An alternative way could be to extrapolate quantum calculations with a Neural Network which is supposed to keep in memory rules given by quantum calculations.
The problem we are trying to solve is finding the minimum energy of the interaction between ice and Pd. This problem is directly related to the size of the cell used, in that sense that making a calculation with a small cell, restricts the number of possible constructions ; it is easy to understand this idea by thinking that more atoms we have, more combinations we can do. Then, in order to find a particular state, we have to use the biggest cell that we can. Rather than calculate every systems, we can only calculate some of them and extrapolate the others by a Neural Network, taking into account the known configurations. The idea is not really to be able to replace a quantum calculation by a neural network one keeping the same precision, but more to select a kind of small window where we have to look for more carefully to find the right system.

II.RESEARCH ACTIVITIES

A. Introduction
Several studies of the last years about interactions between water and metal surfaces show us the importance of understanding chemical and physical properties at the nano scale. These interactions are involved in many processes such as corrosion, catalysis, electrochemistry, material science and many others. Despite the time and energy devoted by a lot of scientists and because of many problems related to the time of calculation often infinitely long, a lot of open questions still remain unsolved.
Indeed, ab-initio calculations and more particulary Density Functional Theory (DFT) are very time-consuming. In order to save time, scientists are obliged to consider small sized systems. Nevertheless it appears obvious that by this restriction, many properties and studies can not be investigated.
My work focuses on the investigation of water on top of metal surfaces with ab-initio calculations, and trying to find an efficient way to calculate large systems.

B. Objectives and Methods
We are investigating, the interaction between one layer of ice on top a palladium surface (111). we are interested in the comparison between different Exchange and Correlation functionals in order to be able to reproduce the experimental datas of the coupling water-metal. Indeed, the typical functionals used to perform calculation with pure metal or water have a completely different form. This is not straight forward to have a functional which can also describe the localized effects of water and the delocalized ones due to metal.
According to ice rule and considering the smallest cell (1x1), there are only two different possible configurations of one layer of ice on top of a metal surface. A cell bigger than (1x1) obviously contains more possible configurations for ice. We counted, for example that there are 2400 different configurations for a cell (2x2). It is impossible to investigate, by means using DFT every configurations of these kind of systems . An idea to solve this problem is to use Neural Networks (NN).
Based on the properties of the few studied configurations, we then extrapolate the results to the unknown points via the NN. (following the model : Perceptron Multi Layers).