The power of modern computers makes it possible to perform simulations of highly complex chemical systems. These simulations provide information on the motions of thousands or sometimes even millions of atoms, which can be very difficult to interpret. A commonly employed analysis strategy is to calculate a histogram along a few collective variables (CVs) as from this one can calculate the underlying free energy. Finding appropriate collective variables is difficult and is often done by using physical/chemical intuition obtained from experiments. This is obviously problematic if we want to predict new chemical structures or novel reaction mechanisms based on simulations alone. We are thus interested in using machine learning algorithms to generate simplified representations of the data obtainable from atomistic simulations so that it can be more easily understood and interpreted by a human user.
The work in the department on these lines focusses on the application and the development of the sketch map algorithm [1, 2]. This algorithm takes as input a set of high-dimensional landmark frames and generates a low-dimensional map that shows how these points are distributed in the high-dimensional space. Projections of further high-dimensional points can then easily calculated using the data on the positions of the landmarks and their projections. We have successfully applied this algorithm to studying problems in protein folding . The sketch-map code is available to download from http://sketchmap.berlios.de.
 A Ardevol, F Palazzesi, GA Tribello, M Parrinello. Journal of chemical theory and computation 12 pp 29-25 (2015)
 A Ardevol, GA Tribello, M Ceriotti, M Parrinello. Journal of chemical theory and computation 11 pp 1086-1093 (2015)
 M. Ceriotti, G. A. Tribello and M. Parrinello. Proceedings of the National Academy of Sciences, (2011).
 G. A. Tribello, M. Ceriotti and M. Parrinello. Proceedings of the National Academy of Sciences, 109(14), 5196–5201 (2012).