INVESTIGADORES
MILLÁN RaÚl Daniel
congresos y reuniones científicas
Título:
Harnessing manifold learning in computational mechanics: collective variables in molecular dynamics
Autor/es:
DANIEL MILLÁN; BEHROOZ HASHEMIAN; MARINO ARROYO
Lugar:
Blois Castle
Reunión:
Workshop; 2nd International Workshop on Reduced Basis, POD and PGD model Reduction Techniques; 2013
Institución organizadora:
Ecole Centrale Nantes
Resumen:
Extracting thermodynamic information such as free energy from molecular
dynamics simulations for biomolecules is very challenging, as the sampling is seriously impaired by a huge
configuration space with many metastable basins. A myriad of method have been proposed to overcome this
difficulty, for instance by enhancing the sampling and the phase space exploration, or by focusing on transition
paths between specific conformations. A family of successful methods relies on the identification of reaction
coordinates, and applying biases along them to overcome the energy barriers. We propose a method that
automatically identifies low dimensional smooth and nonlinear data-driven collective variables (CVs) on the
basis of MD trajectories, or other ensembles characterizing the large deviation flexibility of the molecule, from
the output of nonlinear manifold learning algorithms. These smooth data-driven (as opposed to insight-driven)
CVs provide a mechanistic description of the conformational transitions, and are the basis of accelerated MD.