Mini colloquiums abstract > Materials modelling: from electronic structure to machine learning

Modeling is an essential tool for designing and understanding the behavior of complex technological materials, especially when their conditions of use make experimentation difficult, as in the nuclear field for example. The approaches implemented are both multi-model (physics, chemistry, mechanics, etc.) and multi-scale. Since we want to understand the phenomena and behaviors of materials on a macroscopic scale, and we want to predict the evolution of their properties, we need to start with data on a microscopic scale and work backwards to mesoscopic and macroscopic scales. This means mastering the description of systems at the different scales involved, being able to identify the relevant parameters and link them together, and making the different modules relating to different spatial, temporal or thematic domains communicate with each other (coupling or chaining), paying particular attention to the quantification and propagation of global errors/uncertainties linked to the input parameters, the models used, the numerical implementation or the accuracy of the reference data. In this context, the aim of this mini-colloquium is to bring together theoretical researchers whose common denominator is to model realistic materials, i.e. under conditions close to experimental observations, or even conditions and properties of use (applications), in order to review with them the methodological and numerical advances made in the various approaches at different scales (atomic, mesoscopic, macroscopic), with a particular focus on the need to mix different methods combining electronic structure (DFT, Strong Bonds) and statistical physics (Monte Carlo, Molecular Dynamics) and the problem of scale change. Particular emphasis will be placed on probabilistic mathematical methods for quantifying uncertainties in numerical codes involving a very large number of variables or parameters. Finally, this mini-colloquium will compare these multi-scale, multi-physical approaches with Artificial Intelligence (AI) methods, and more specifically with Machine Learning (ML) methods.