Atomistic simulations based on an atomistic description of matter have become a most important scientific tool of contemporary material science investigations. Yet much remains to be done to extend the scope of these simulations. 

Very often, the quality of the models used falls short of what would be needed; the size of the system that can be simulated is too small or the-time scales that can be simulated much too short to be of relevance. 

We show how modern machine learning techniques, working in a synergetic way with molecular dynamics simulation, offer a way of overcoming these problems to bridge the gap between simulation and real life experiments.