Ab initio crystal structure prediction is well known to have high computational cost resulting from the exploration of a huge phase space with many degrees of freedom. Despite intelligent strategies implemented for the search, much of the time is still spent in performing ab-initio calculations in regions of the phase space far from the global minimum.
In this work we propose a new strategy based on ab-initio quality energetics obtained from deep neural-networks as implemented in the TensorFlow library to increase the efficiency of the search and demonstrate its effectiveness in the example of Glycine molecular crystals. We start with modified Behler-Parrinello symmetry functions to build single-atom environment vectors to represent the crystal structure. We train a layered, all-to-all connected network for each atomic type on a database of ab-initio calculations. The total energy is expressed as a sum of individual, environment dependent, atomic contributions. We explore the machine learning performance of such a network not only for energies but also for forces and pressure with respect to training data, network architecture and training process. We then show the harmonious integration of this strategy into the existing crystal structure prediction workflow and test its applicability to system sizes that could not be tackled otherwise. References: Lyakhov A.O., Oganov A.R., Stokes H.T., Zhu Q. (2013). Comp. Phys. Comm. 184, 1172-1182  tensorflow.org  J.Behler and M. Parrinello (2007). Phys.Rev.Lett. 98, 146401  J. S. Smith, O. Isayev, A. E. Roitberg, (2017), Chem. Sci., 8, 3192-3203