High Performance Simulations across Computer Architectures
Tuesday, 27 March 2018
11:00am - 12:30pm
Campus Stuttgart Vaihingen, Pfaffenwaldring 9
Intermolecular potentials from machine learning across chemical space
Tristan Bereau, Max Planck Institute for Polymer Research
Classical intermolecular potentials typically require an extensive parametrization procedure for any new compound considered. To do away with prior parametrization, I will describe a combination of physics-based potentials with machine learning that is transferable across small neutral organic molecules made of H, C, N, and O atoms. Unlike other potentials, the model is transferable in its ability to handle new molecules and conformations without explicit prior parametrization. I will describe applications on gas-phase dimers and discuss perspectives toward the condensed phase.
Machine learning potentails for molecular dynamics
Philipp Marquetand, University of Vienna
Machine learning is used to find the relationship between the nuclear geometry of a molecule and the corresponding properties, like the potential energy or the dipole moment. In this way, the machine learning algorithm – an artificial neural network in this case – serves as a highly accurate and extremely fast tool during the simulation of molecular dynamics, i.e., the motion and behavior of a molecule in time. Different developments will be presented, e.g., how the neural networks learn to assemble small, known molecules serving as building blocks to predict the properties of a large unknown molecule. As an application, calculations of infrared spectra obtained via molecular dynamics simulations will be shown.
Minimally Invasive Integration of Tree-Structured Cartesian Grids in Existing Applications
Michael Lahnert, University of Stuttgart
We present our approach for minimally invasive integration of dynamically-adaptive tree-structured grids at the example of the molecular dynamics (MD) simulation code ESPResSo. It includes an implementation of the Lattice-Boltzmann method (LBM) to subject a molecular ensemble to a background flow as well as a continuous representation of the electrokinetic equations. We port the grids of ESPResSo's physical subsystems to an extended version of p4est, a well-known and scalable library for tree-structured Cartesian grids. Our contribution to p4est allows simpler integration into existing applications. As not all grids are discretized in the same way, we use independent p4est instances for each algorithm and describe our approach for reducing communication in the coupling scheme. We show results and scalability tests for the individual components as well as for the integrated application on different hardware architectures.
Interpolation of Potential Energy Surfaces using Gaussian Process Regression
Alexander Denzel, University of Stuttgart
In this work we are combining methods of quantum mechanics with machine learning methods to find interesting geometries of molecules and other useful properties of chemical systems. One of the fundamental problems in theoretical chemistry is the steep scaling of the calculations that are necessary to determine the energy and its derivatives in a chemical system. We use machine learning methods like neural networks and Gaussian process regression (GPR) to interpolate between obtained energies. GPR is easy and fast to set up for a new system. Furthermore, formulated in a Bayesian setting, it has the capability of giving uncertainty measurements and the possibility of optimizing its parameters with e.g. the maximum likelihood method. We apply this method to interpolate several energy surfaces to do geometry optimization and reaction rate calculations that incorporate quantum mechanical tunneling with fewer electronic structure calculations and therefore, faster than before.