Title: 

" Development of Machine Learning Formalisms for Atomic Diffusion and Electron Densities” 

ABSTRACT: 

Predicting diffusion of ions and atoms using computational chemistry can facilitate the design of advanced materials with targeted transport properties. One way to observe and characterize particle diffusion is by running long molecular dynamics (MD) simulations. Methods like density functional theory (DFT) can be used to perform such simulations, however, they are limited to small system sizes (hundreds of atoms) and short timescales (tens of picoseconds). To overcome these computational limitations, machine learning (ML) potentials can be developed in order to study diffusion of large-scale systems (many thousands of atoms) for longer times (many nanoseconds). These ML potentials are generally trained using data generated from DFT by exploring a large potential energy surface for a give system.  

Design and characterization of anhydrous proton conduction materials are essential for next generation proton exchange membrane (PEM) fuel cells, allowing operation at elevated temperatures and low humidity. Graphanol (hydroxy functionalized graphane) is a material that was previously shown by DFT calculations to enable anhydrous proton diffusion. However, calculating the proton diffusion barrier and understanding the proton hopping mechanism for a fully functionalized graphanol system proved to be impossible using DFT. We used the DeePMD formalism with active learning (deep generator) to construct robust deep potentials (DP) for graphanol. These DPs were tested and found to be highly accurate to study proton dynamics and bond energetics of the system. We estimated the proton diffusion barrier of graphanol to be 93 meV, which is lower than conventional PEM membranes like Nafion. There are other studies that we will be performing using these accurate DPs, such as: (1) to fully characterize diffusion mechanisms; (2) accounting for defects; (3) identifying mechanisms for hopping from one graphanol sheet to another; (4) influence of hydration on proton conduction. Tracking protons like classical particles does not account for electron densities and polarization effects. Thus, we are working on a DeepCDP (Deep Learning Charge Density Prediction), which is aimed at predicting electron densities for infinitely large system sizes. The diffusion of charges will be tracked based on the movement of the charge centers. We also showed that it is possible to use accurate DPs for metal-organic frameworks (MOFs) like UiO-66 in concert with classical potentials for adsorbates to accurately compute diffusivities through a hybrid potential approach. This hybrid approach was used to model the diffusion of Ne and Xe through UiO-66. We plan to use this approach to test graphanol in the presence of guest fluids like water and observe its influence on proton conduction. ML potentials were also used to model interface diffusion of chalcogenide alloys and electrodes that are used in next generation non-volatile memory cells.  These potentials were constructed using the moment tensor approach (MTP). MTP-based active learning helped in exploring amorphous phases of these materials. This methodology will be used to find advanced chalcogenide alloys – electrode interfaces that are non-diffusive for long MD time.