Interpreting total X-ray scattering data and lithium battery cycling curves using atomistic simulation and machine learning

Pair distribution function (PDF) is a total scattering technique for determining the local structure of nanoscaled materials. In this thesis, we combine atomistic simulations and machine learning models to (i) understand the point defect in ceramics oxide with experimentally-measured PDFs, and (ii) to investigate the lithium occupancy during Li battery charging process using in-situ PDFs. We also developed a data-driven model for lithium battery cycling curve prediction. This thesis paves way for future studies to further explore the possibility of computational modelling and data-driven method aided material characterization and discovery. 

A workflow is presented for performing PDF analysis of defected materials using structures generated from atomistic simulations. A large collection of structures, which differ in the types and concentrations of defects present, are obtained through energy minimization with an empirical interatomic potential. Each of the structures is refined against an experimental PDF. The structures with the lowest goodness of fit R_w values are taken as being representative of the experimental structure. The workflow is applied to anatase titanium dioxide (a-TiO₂) and tetragonal zirconium dioxide (t-ZrO₂) synthesized in the presence of microwave radiation, a low temperature process that generates disorder. The results suggest that titanium vacancies and interstitials are the dominant defects in a-TiO₂, while oxygen vacancies dominate in t-ZrO₂. 

To accelerate the process of conventional PDF analysis, feature extraction and a neural network model are applied to reach the similar goal with PDF data. A dataset of TiO₂ structures with vacancies and interstitials of oxygen and titanium is built and the structures are relaxed using energy minimization. The features of the calculated PDF of these defected structures are extracted using linear methods and nonlinear methods. The extracted features are used as the inputs to a neural network that maps the feature weights to the concentration of each defect type. A physics-based initialization of the autoencoder has the highest accuracy in predicting the defect concentrations.

In-situ PDF measurement during Li battery charging process was used to determine the Li occupancy of Li_xTiO₂(B) phase. The occupancy at x = 0.25 and x = 1.0 was first verified against literature results with MD simulations and PDF analysis. For 0.25 < x < 1.0, a large amount of structures were created with varying Li concentrations and occupancy and PDF analysis was performed. It is found that C site is the most favorable when x = 0.25, A1 and A2 site is the most favorable when x = 1.0, when 0.25 < x < 1.0, the C sites becomes less favorable and the Li ions starts to fill in A1, and A2 sites, with A2 sites filled faster than A1. 

Regarding monitoring the health condition of Li battery, we developed a data-driven model to predict the entire constant-current cycling curve with limited input information that can be collected in a short period of time. We collected a total of 10,066 charge curves from 52 batteries. We applied feature extractions and a multiple linear regression model to predict a complete cycling curve based on a limited portion of it. We demonstrate that both single continuous segment and multiple separated segments can be used as the input to predict the complete cycling curve. The model achieves a 2% prediction error of an entire charge curve using only 10% of the curve as the input for the LiNiO₂-based batteries and achieves the same accuracy with only 5% of the curve as the input for the LiCoO₂-based batteries.