Atomistic computer simulation has established itself as one of the most important methodologies in modern science, its impact being felt in areas as diverse as physics, chemistry, geophysics, materials science and biophysics, to name but a few. Yet in spite of great progress in computer power and algorithms, we are still not able to simulate such important phenomena as nucleation, phase transitions or protein folding. The purpose of this proposal is to strongly push back the limits of length, time scale and accuracy of present-day methods, greatly enhancing the scope of atomistic simulations. We expect the impact of a successful outcome of this proposal to revolutionize the field. We shall make use of three technical innovations: an extension of Langevin-type equations to include correlated (colored) noise; the use of h-matrices to speed up electronic structure calculations; and an intelligent use of neural networks. Our strategy will be complex. We plan to speed up ab-initio molecular dynamics calculations considerably and also to generate new and highly accurate effective potentials based on electronic structure calculations. A large part of our effort will be devoted to the time scale problem. In this respect we shall improve metadynamics so that its implementation becomes as general and as automatic as possible, and we shall also introduce methods for reconstructing the real dynamics from metadynamics. Finally, highly innovative and powerful sampling methods based on specially designed colored noise Langevin equations will be developed.