Simulations improve the theoretical description of experiments probing elusive properties of water

Scientists from EPF Lausanne have improved the theoretical description of complex non-linear optical measurements by developing an atomic-scale model of the underlying physical process. The researchers hope that this model, which was developed by combining high-end quantum chemistry methods, machine learning and large-scale calculations using “Piz Daint”, will provide better insight into experiments that investigate the puzzling behavior of water. 

October 3, 2017 - by CSCS

Water is a key substance in Chemistry, Biology and Materials Science, an active medium within which cellular process occur and technologically crucial reactions are performed. The reason behind its ubiquity is the formation of strong, directed interactions (hydrogen bonds) between neighboring molecules, resulting in extended, complex networks. In order to probe these networks experimentally, techniques are required that are sensitive to the mutual orientation of molecules.

SHS measurements are uniquely suited to probe the complex behavior of water. The group of the Laboratory For Fundamental Biophotonics (LBP) of Sylvie Roke, professor at EPF Lausanne has pioneered the application of these methods to pure water, salt solutions, membranes and even living cells. Interpretation of these complex experiments, however, relies heavily on theoretical modelling. In the past, a number of approximations have been made in analyzing SHS experiments. Key among these was that the hyperpolarizability of a molecule does not depend on the surrounding environment.

To test this assumption, members of the Laboratory of Computational Science and Modelling (COSMO) group of Michele Ceriotti, also professor at EPF Lausanne, performed large-scale molecular dynamics calculations with simulations boxes containing almost 300’000 water molecules, that were made possible thanks to the high-performance, GPU-accelerated computing capabilities of “Piz-Daint” at CSCS. From the output of these simulations, 10,000 molecules were chosen (so as to sample a variety of different environments) and their hyperpolarizabilites computed using high-level quantum-mechanical calculations. The resulting values varied greatly between different environments, showing clearly that this common assumption has to be reconsidered.

Machine-learning model to predict hyperpolarizability

Finally, the researchers developed a machine-learning model to predict the hyperpolarizability of water molecules based on a small number of “training” calculations, that proved to be nearly as accurate as the reference quantum simulations. The performance of this model promises to provide quantitative accuracy when computing experimental SHS intensities, making it possible to obtain insight into the behavior of water in more complex, biologically and technologically relevant scenarios.