Prof. Michele Parrinello
Department of Chemistry and Applied Biosciences, ETH Zurich, Switzerland,
Facoltà di Informatica, Istituto di Scienze Computazionali, Università della Svizzera italiana, Lugano, Switzerland,
Istituto Italiano di Tecnologia, Genoa, Italy
Atom based computer simulation is one of the most important tools of contemporary physical chemistry. In spite of its many successes, it suffers from severe limitations. Here we show how machine-learning techniques can help in solving at least two different problems. The first one is the accuracy of current interatomic potential models; the second is the limited time scale that simulations can explore. In order to solve the first problem we train a neural network on a set of accurate but expensive quantum chemical calculations. In this way, it is possible to obtain an accurate description of the system at a relatively low computational cost. Crucial for the success of this program has been the design of the neural work and the selection of the training set. We apply this approach to study a metal non-metal transition and to chemical reactions in condensed phases. These applications would not have been possible without the use of efficient sampling methods capable of lifting the time scale barrier. To this effect, we have developed two very efficient sampling methods, metadynamics and variationally enhanced sampling. Both methods are based on the identification of appropriate collective variables, or slow modes, whose sampling needs to be accelerated. Machine learning can be used also for the construction of efficient collective variables based on a modification of the well-known linear discriminant analysis classification method. Finally, we use the variational enhanced sampling approach and a deep neural network to further increase our sampling ability.