Deep generative models for inverse design of sodium-ion cathode materials

Lithium-Ion Batteries (LIBs) are the most commercialized secondary batteries for large-scale Energy Storage Systems (ESS), yet they rely on scarce and costly materials. Sodium-Ion Batteries (SIBs) are considered a promising alternative due to the abundance and low cost of sodium, but their performance has not yet matched that of LIBs. Improving SIBs requires the discovery of optimal cathode materials that combine high capacity, stability, and ionic conductivity. Among these, polyanion sodium cathodes are especially attractive for their structural stability and high voltage, though there remains a need for strategies that can further boost their electronic and ionic conductivity. This thesis presents five studies aimed at identifying
optimal configurations of polyanion sodium cathode materials.

First, a Density Functional Theory (DFT) screening was performed on the unit cells of four polyanion sodium-ion cathode systems: NaTMPO₄ (olivine), NaTMPO₄ (maricite), Na₂TMSiO₄, and Na_2.56TM_1.72(SO₄)₃, with transition metals TM ∈ {Fe, Mn, Co, Ni}. The screening considered both single- and multi-TM ion compositions within the unit cell. To capture ionic conductivity, Molecular Dynamics (MD) simulations were carried out using a Machine Learning Interatomic Potential (MLIP) in an active learning framework, enabling the exploration of ion mobility over extended time and length scales for the single-TM cathode materials. This study resulted in a dataset that provides a diverse representation of the polyanion sodium cathode material phase space and serve as a robust foundation for training MLIPs.

Second, the Open-Circuit Voltage (OCV) of NaFePO₄, determined from the DFT screening, was compared to the experimental voltage profile. This analysis revealed that unit cell exploration alone cannot reproduce experimental observables. While first-principle methods are essential to describe possible phase transitions accurately, they remain computationally demanding for large-scale systems, highlighting the need for computationally efficient approaches with accuracy comparable to first-principles calculations.

Third, state-of-the-art universal MLIPs were compared with a system-specific MLIP trained on the polyanion sodium cathode materials dataset from the first study. Across three key simulation tasks — structural optimization, MD simulations, and Nudged Elastic Band (NEB) calculations — the system-specific charge-PaiNN (cPaiNN) model performed better than the universal MLIPs, while introducing finetuning of the universal MLIPs made them equal to the system-specific cPaiNN. A key advantage of cPaiNN is its explicit atomic charge descriptor, enabling accurate property prediction and direct insight into TM ion redox processes, while maintaining first-principles accuracy with greater computational efficiency and interpretability.

Fourth, Cluster Expansion (CE) models trained on DFT- and MLIPbased datasets were used together with Monte Carlo (MC) sampling to predict thermodynamically stable configurations in Na_xMn_yFe_{1 – y}PO₄ (olivine). However, the CE predictions did not match experimental observables, whereas direct DFT and MLIP calculations correctly identified the experimental structure as lowest in energy. This highlights the need for alternative methods to efficiently sample stable configurations across different charge states of the cathode materials.

Fifth and finally, Dis-GEN is introduced as a generative model for site-disordered inorganic crystal structures. It incorporates both compositional disorder and vacancies while preserving crystallographic symmetry, enabling systematic exploration of site-disordered functional materials. While Dis-GEN can identify promising dopants for polyanion sodium cathode materials, integration with MLIPs, or other external methods, is required to efficiently determine thermodynamically stable configurations within the vast and complex phase space of these materials.

The findings of this thesis provide atomic-level insights into polyanion sodium cathode materials and introduce methods and results that enable accelerated discovery of these cathode materials.