DeepIce: A Deep Neural Network Approach To Identify Ice and Water Molecules

Abstract

Computer simulation studies of multi-phase systems rely on the accurate identification of local molecular structures and arrangements in order to extract useful insights. Local order parameters, such as Steinhardt parameters, are widely used for this identification task; however, the parameters are often tailored to specific local structural geometries and generalize poorly to new structures and distorted or under-coordinated bonding environments. Motivated by the desire to simplify the process and improve the accuracy, we introduce DeepIce, a novel deep neural network designed to identify ice and water molecules, which can be generalized to new structures where multiple bonding environments are present. DeepIce demonstrates that the characteristics of a crystalline or liquid molecule can be classified using as input simply the Cartesian coordinates of the nearest neighbors without compromising the accuracy. The network is flexible and capable of inferring rotational invariance, and produces a high predictive accuracy compared to the Steinhardt approach, the tetrahedral order parameter and polyhedral template matching in the detection of the phase of molecules in premelted ice surfaces.

A generalized deep learning approach for local structure identification in molecular simulations

“…We achieved similar overall accuracy (∼95%) as reported in the original work 22. An approach that was published during the writing of this paper achieved higher classification accuracy but only distinguished between the ice Ih and liquid phases 26…”

“…The parameters for the radial pooling functions were learned during training. The DeepIce26 network of Fulford et al consisted of four subnetworks—Cartesian coordinates network, spherical coordinates network, Fourier transform network, and spherical harmonics network. The coordinates-based subnetworks used the Cartesian and spherical coordinates of neighbors with respect to the central atom as input, respectively.…”

“…Others have used machine learning to develop force fields,14,15 for coarse-graining,16,17 to identify reaction coordinates,18 and to extract trends in results by clustering similar structures 19–21. There are a few previous efforts to use machine learning for structure identification in molecular simulations 19,22–26. In general, such approaches seem promising given the widespread success of machine learning in tasks such as computer vision ( e.g.…”

“…Similar symmetry functions have been used in combination with support vector machines to predict particle rearrangements under shear30 and at grain boundaries 31. Other approaches for structure identification in molecular simulation have first processed the simulation output with spherical harmonics24,26 or other carefully engineered features 25,26. Panagiotopoulos and co-workers took a different approach, using neighborhood graphs and diffusion maps32,33 to discover, cluster, and classify crystal structures and identify relationships between crystal types without explicit training of each structure 19,23.…”

Simulating solvation and acidity in complex mixtures with first-principles accuracy: the case of CH$_3$SO$_3$H and H$_2$O$_2$ in phenol

“…Fortunately, data-driven techniques provide a framework that can also be used for unbiased structural characterization. [48][49][50][51][52][53][54] Here we use the sketch-map dimensionality reduction algorithm. [55][56][57] Similar to multi-dimensional scaling, 58 sketch-map tries to find a low-dimensional representation of a set of configurations, matching the distances between high-dimensional sets of features that describe each structure, and those between their projections.…”