Practical use of information concepts in molecular simulation
Gabor Csanyi, University of Cambridge
In this talk I will show, through three examples, how simple concepts from applied probability and machine learning have the potential to change the art of atomic and molecular scale modelling. (i) Bayesian inference can be used to ensure the sound treatment of information, thus maximising the utility of simulations, for example in reconstructing free energy surfaces from molecular dynamics. (ii) In constructing interatomic potential models, Gaussian processes are used to encode prior information and lead to an unprecedented combination of accuracy and computational efficiency - as demonstrated by the most accurate model of water to date that correctly captures the behaviours of clusters, ice and liquid phases. (iii) a fresh look at the sampling problem in the context of exploring potential energy landscapes yields a gain of many orders of magnitude over 'replica exchange', and brings into clear focus one of the central problems in molecular simulation: designing a process to minimise the decorrelation time or - equivalently - to accelerate the convergence to any specified invariant distribution.