Machine Learning and Crystal Structure Prediction of Molecular Crystals

There is a natural synergy between data-hungry machine learning methods and crystal structure prediction of molecular crystals that requires a careful search in a vast potential energy landscape. In our previous study we demonstrated how taking advantage of machine learning methods can enable novel predictions even for well-studied molecular crystals [1]. In order to leverage this synergy further, we have been developing a deep neural network training tool, PANNA (Potentials from Artificial Neural Network Architectures), based on TensorFlow framework [2]. In creating transferable machine-learned potentials, the key step is the non-linear process of training the network model. We will demonstrate a variety of network training techniques that can be explored within PANNA, from ones that are commonly used in machine learning community to the ones that are specific to atomistic simulations. We will report the effect of data selection, input representation and training methods on the training dynamics and on the resulting potentials in the difficult case of molecular crystals.

[1] C. Bull et al. “ζ-Glycine: insight into the mechanism of a polymorphic phase transition” IUCrJ 4, p.569 (2017)
[2] M. Abadi et al. “TensorFlow: Large-scale machine learning on heterogeneous systems” (2015)