Ceci est une ancienne révision du document !

Séminaires/Colloquium 2019

Machine learning in atomistic simulations: from reaction pathways to phase diagrams

05 april 2019 à 9h00
COLLOQUIUM Centre Blaise Pascal dans le cadre du E-CAM, Extended Software Development Workshop : Topics in Classical MD Lieu : Salle 1 place de l'École, ENS Lyon, France
Christoph DELLAGO , (Faculty of Physics, University of Vienna)

Abstract :
Atomistic computer simulations of processes occurring in condensed matter systems are challenging for several distinct but related reasons. For large systems, the accurate calculation of energies and forces needed in molecular dynamics simulations may be computationally demanding, particularly if electronic structure calculations are used for this purpose. Other difficulties arising in the dynamical simulation of condensed matter processes consist in detecting local structures characteristic for stable or metastable phases and in identifying important degrees of freedom that capture the essential physics of the process under study. In this talk, I will discuss how these problems can be addressed using machine learning approaches. I particular, I will focus on a computational study of water and ice based on a high-dimensional neural network potential trained with ab initio reference data. We have shown that Waals interactions are crucial for the formation of water's density maximum and its negative volume of melting. Our simulations have also revealed that nuclear quantum effects play an important role in modulating the thermodynamics stabilities of different phases of water.

  • T. Morawietz, A. Singraber, C. Dellago, and J. Behler, “How Van der Waals interactions determine the unique properties of water“, Proc. Natl. Acad. Sci. USA 113, 8368-8373 (2016).
  • B. Cheng, E. A. Engel, J. Behler, C. Dellago, and M. Ceriotti, “Ab initio thermodynamics of liquid and solid water”, Proc. Natl. Acad. Sci. USA 116, 1110 (2019).

Fast neural solvers

04 mars 2019 de 11h00 à 12h00
COLLOQUIUM Centre Blaise Pascal-Laboratoire de Physique
Amphi. Schrödinger, ENS Lyon, France

Patrick PEREZ , (Valeo AI)

Organisateurs :

  • Corentin Herbert (Laboratoire de Physique, ENS de Lyon)
  • Cerasela Calugaru (Centre Blaise Pascal, ENS de Lyon, France)

Abstract :
Modern artificial neural networks dominate a number of classic machine learning tasks in a wide range of application domains. What is probably less known is that they also offer new ways to attack certain optimization problems, such as inverse problems arising in physics or image processing. While a variety of powerful iterative solvers usually exist for such problems, deep learning may offer an appealing alternative: With or without supervision, neural networks can be trained to produce approximate solutions, possibly of lower quality, but orders of magnitude faster and with no need for initialization. We shall discuss different ways to design and train such fast neural solvers, with examples from computer vision and graphics.

animation/seminaires/2019/accueil.1553272563.txt.gz · Dernière modification: 2019/03/22 17:36 par cicaluga