Optimization of Energy Storage Materials through Ab-initio Calculations and Deep Learning Techniques in Material Science

Abstract

The growing demand for high-energy-density and long-cycle-life lithium-based batteries has intensified the need to understand and optimize the interfacial phenomena that govern electrochemical stability, ion transport, cycling performance, and safety. Among these interfacial phenomena, the solid electrolyte interphase (SEI) is of central importance as it mediates Li-ion transport while passivating the electrode surface and suppressing continuous electrolyte decomposition. The SEI is a thin, complex layer formed through electrolyte decomposition at the electrode–electrolyte interface. Despite its critical role, the SEI remains one of the least understood components of electrochemical cells because of its chemical heterogeneity, structural complexity, and dynamically evolving nature. These characteristics make direct experimental characterization of Li-ion transport mechanisms within the SEI especially challenging. To address this challenge, this thesis combines first-principles calculations and machine learning approaches to investigate and model Li-ion diffusion processes in lithium-based energy-storage materials, with particular emphasis on Li-ion transport through the inorganic sublayer of the SEI. The first part of this work focuses on the development of a generalized machine learning framework, named the Machine Learned Diffusion Coefficient Estimator (ML-DiCE), for predictive modeling of diffusion coefficients in multicomponent materials. A large experimental diffusion-coefficient dataset covering impurity, self-, and chemical-diffusion mechanisms in impure metallic and multicomponent alloy systems was curated and used to train Random Forest, Deep Neural Network, and Support Vector Regression models. Through composition-aware feature engineering and systematic hyperparameter optimization, the developed models achieved strong predictive performance, demonstrating the capability of data-driven approaches to capture complex atomic diffusion behavior across diverse material systems. This work represents one of the first systematic efforts to model diffusion coefficients in complex solid-state material systems using machine learning techniques. The predictive framework and physical insights developed from this study provide a foundation for extending data-driven diffusion modeling toward Li-ion transport in SEI materials. The second part of this thesis investigates Li-ion diffusion in inorganic SEI components using density functional theory (DFT) combined with deep graph neural network models. A comprehensive diffusion-coefficient dataset was generated using nudged elastic band (NEB) calculations for eight major inorganic SEI components: LiF, LiCl, LiBr, LiI, Li2O, Li2S, Li3N, and Li2CO3. The dataset includes both grain and grain-boundary diffusion pathways across homogeneous and heterogeneous interfaces. A path-aware graph variational autoencoder (GVAE) was employed to learn latent representations of diffusion trajectories, which were subsequently integrated into graph neural network (GNN) models for potential-terrain-aware migration-barrier prediction. The combined GVAE–GNN framework achieved strong predictive performance, with test-set $R^2$ values of 0.93 and 0.94 for grain and grain-boundary diffusion, respectively. The training strategy enabled the model to learn features associated with the potential-energy terrain governing Li-ion diffusion in both grain and grain-boundary systems. The analysis revealed that Li-ion migration is governed not only by SEI chemistry but also by local microstructural features such as exposed surface orientation, grain-boundary character, and interfacial mismatch. These results establish a clear relationship between the structure of inorganic SEI components and Li-ion transport. It also suggest that suppressing dendrite growth may require uniform Li-ion migration barriers and evenly distributed migration channels across the inorganic sublayer of the SEI. Achieving this uniformity depends on careful control of both phase composition and interfacial microstructure in SEI design. Finally, this thesis explores MXene-assisted interfacial engineering as a case study for stabilizing lithium-metal/SEI interfaces and regulating Li-ion flux. Using combined DFT calculations and numerical descriptor analysis, Li-metal/MXene heterostructures with different surface terminations were systematically screened. The study demonstrates that mechanically optimized and electrochemically compatible MXene coatings can form stable interfaces with lithium metal while promoting uniform ion transport and minimizing interfacial distortion. In particular, ABC-stacked, Cr-based, fluorine-terminated MXenes exhibited favorable adsorption energetics and enhanced interfacial stability, indicating their potential as auxiliary SEI layers for Li-metal electrodes to suppress dendrite growth and improve long-term electrochemical performance. Overall, this thesis establishes integrated computational frameworks that combine atomistic simulations, diffusion-coefficient modeling, and machine learning to accelerate the understanding of ionic transport phenomena in SEI materials and guide the rational design of stable SEI. The methodologies and insights developed in this work provide a foundation for engineering SEI architectures through controlled ion transport, optimized interfacial chemistry, and microstructural regulation, thereby enabling the design of high-performance interfaces for next-generation lithium-based energy-storage systems.