(360x) Studying Anhydrous Proton Transport on Graphene-Based Materials Using Deep Learning Methods | AIChE

(360x) Studying Anhydrous Proton Transport on Graphene-Based Materials Using Deep Learning Methods


Achar, S. - Presenter, University of Pittsburgh
Bernasconi, L., University of Pittsburgh
Zhang, L., Princeton University
Johnson, K., University of Pittsburgh
The unique physical and chemical properties of graphene-based materials have spurred research interests in fuel cell applications, especially as proton exchange membranes (PEM). The design and study of such new proton conducting materials are essential for developing PEM fuel cells that operate at intermediate temperatures and conditions of low humidity. We show that graphane (hydrogenated graphene) functionalized with hydroxyl groups (graphanol) is a very promising material for anhydrous proton conduction. Previous methods to study proton conduction on graphanol involved performing computationally demanding density functional theory (DFT) calculations, limited to small systems and short time-scales. To overcome this computational hurdle, we developed dense neural network potentials using deep learning. Our previous work with using deep-learning potentials (DP) for graphane (the precursor material of graphanol) showed tremendous capabilities for capturing molecular physics with small training data sets. We have developed a DP for graphanol that is capable of accurately simulating both charged and uncharged systems. This DP predicts DFT forces to within meV/Å accuracy. We show that our DP is capable of accurately predicting vibrational and thermal properties, dynamic properties, and proton mobility of graphanol. We discuss active learning strategies required to achieve high accuracy for the graphanol DP.

Proton conductivity involves charge transfer and therefore cannot be fully described by observing motion of atoms alone. However, our graphanol DP potential does not include any information about electron density, charge polarization, or transport of charge centers. We have developed a deep learning formalism for predicting charge densities only using atomic positions for molecules and periodic systems. Our method, called DeepCDP (Deep-learning Charge Density Prediction), can predict electron densities for arbitrarily large systems. A DeepCDP model is trained on electron densities generated from DFT-MD simulations. Smooth Overlap of Atomic Orbitals (SOAP) is used to fingerprint atomic environments to their corresponding electron densities on a grid-point basis. Our DeepCDP model for graphanol achieves electron density prediction accuracies of >99% for both charged and uncharged systems. We demonstrate methods to use these predicted electron densities to compute the diffusion of the center of charge from DP simulations of graphanol with excess protons. Our combination of DeePMD and DeepCDP provides a complete description of proton conductivity solely using deep learning methods.