The demand for efficient energy-storage systems is rapidly growing, together with the need to identify clean energy sources, such as hydrogen. To these purposes, the characterization of materials at the atomistic scale is crucial, since it can provide unique insights to improve their performance, especially in operando conditions. Obtaining an experimental characterization of these systems with atomic resolution is extremely challenging, while ab initio atomistic simulations are, in principle, an ideal tool to investigate dynamical phenomena on the atomic scale. Despite its potential, the prohibitive cost of ab initio molecular dynamics has limited its application to study realistic systems. Recently, machine learning (ML)-based interatomic potentials have emerged as a valuable solution to reconcile the accuracy of ab initio methods with the efficiency of classical force field. These ML potentials are trained to reproduce the energy and forces from a large set of quantum mechanical calculations, and they can be optimized on small system sizes and then used to simulate much larger systems for long (e.g. nanoseconds) timescales. Despite their promise, the construction of these potentials for complex, multicomponent reactive systems represents a complex endeavor, requiring a comprehensive training set that incorporates all relevant configurations. To this end, enhanced sampling methods such as metadynamics allow to speed up the process of collecting configurations and generating uncorrelated structures which cover the complete energy landscape.

-

Guests Lectures: Umberto Raucci - Istituto Italiano di Tecnologia Francesco Mambretti- Istituto Italiano di Tecnologia In this PhD course, we aim to present a comprehensive overview of the current state-of-the-art methodologies in the field. We will review various approaches for training machine learning (ML) potentials and collecting the required training datasets, while highlighting the advantages and limitations of each approach, depending on the specific application. Additionally, the course will feature illustrative applications aimed at deepening participants' understanding of structural dynamics, chemical reaction kinetics, and transport phenomena within energy storage and conversion systems. The course will be structured as follows: 1) ab initio molecular dynamics: general introduction 2) Machine learning interatomic potentials: theoretical foundations and implementation 3) Hand-on section on some case studies The hands-on section will focus on the full training procedure of a ML-potential on a few case studies, using the Quantum Espresso, DeepMD-Kit and LAMMPS softwares.

In presenza

Presentazione orale

P.D.2-2 - Maggio