March, Friday 2nd14:30 (room 2014, 'Digiteo Shannon' 660 building) (see location)
Title: Physicists using and playing with Machine Learning tools: two examples
AbstractIn a first part, I will present a recent work I did on Markov State Modeling. Markov State Modeling has recently emerged a key technique for analyzing molecular simulations. In particular it is used to find the metastable states of complex systems in thermal equilibrium, such as the conformations of a protein undergoing a folding event.
We extend this technique to non-equilibrium problems, and in particular to friction, using a combination of density-based clustering algorithm and RPCCA (Robust Perron Cluster Cluster Analysis). We demonstrate that our method allows the unprejudiced identification of the minimal basis of natural microscopic states necessary for describing the dynamics of sliding, including frictional dissipation and stick-slip events.
In a second part, I will introduce my current work on the automatic identification of defects in supercooled (glassy) liquids. The long-standing problem of identifying the structures responsible for glassiness presents itself as a rather well-posed Machine-Learning problem.
Building on the idea that particles alternate between caged and jumping states, we propose a scheme for classifying each particle's state into jumping or caged, based on its local structure only. Although this scheme already performs well in some respect, I will show that large avenues are left open for future research.
Contact: guillaume.charpiat at inria.fr
All TAU seminars: here