Data-driven graph neural networks guide codoping for stable Ni-rich layered cathodes

Abstract

Ni-rich layered oxide cathode materials have emerged as promising candidates for next-generation mainstream high-energy nonaqueous lithium-based batteries because of their inherent advantages in terms of specific capacity. However, the delicate layered structure is more susceptible to both crystal and morphological structure degradation under electro-chemo-mechanical stresses at high states of delithiation, leading to an unsustainable cycle life and vulnerable safety hazards. To address these issues, 20 extensively utilized doping elements are incorporated into a LiNi0.9Co0.1Mn0.1O2 (NCM90) cathode to generate a critical dataset with 20 descriptors comprising its intrinsic, structural and morphological features. Furthermore, we develop a designed multidimensional graph differential neural network (MDGNN) model trained on the former dataset to fit the doping laws, which results in an error rate in predicting two key performance metrics, e.g. specific capacity and capacity retention, of an arbitrary dopant-modified NCM90 cathode. Combined with the MDGNN model and traditional chemical principle, 1.2 mol% Al3+, 0.3 mol% Zr4+ and 0.5 mol% Sb5+ are elaborately doped into the NCM90 cathode to stabilize its electrochemical and thermal properties, and the related operation mechanism is systematically investigated via characterization. The developed 7.3 Ah Al1.2Zr0.3Sb0.5-NCM90||graphite pouch cell attains 87.3% capacity retention after 3000 cycles at 1.0/1.0 C, verifying the practical feasibility of our MDGNN model. These results open a viable avenue for the data-driven design of high-performance battery materials.