From Physics-Informed Deep Learning to Experimental Realization: Unlocking Entropy-Stabilized Titanium Halide Solid Electrolytes.
Abstract
Accelerating the discovery of high-performance solid electrolytes (SEs) is hindered by scarce experimental data and vast inorganic chemical spaces. Here, we present a physics-informed deep learning platform that integrates a physics-informed neural network with a dual-source graph neural network. By embedding fundamental physical constraints (e.g., lattice stability and coulombic interactions) into the training process, our framework overcomes the small-data dilemma, enabling the reliable screening of over 2 80 000 inorganic crystals. Guided by a multi-objective optimization incorporating ionic conductivity, thermodynamic stability, and global economic factors, the titanium-based halide family is identified as a cost-effective alternative to scarce zirconium- or scandium-based SEs. As a proof-of-concept, a novel Li4TiBr4Cl4 electrolyte is synthesized, which exhibits a high room-temperature ionic conductivity of 1.14 mS/cm. While 0 K density functional theory calculations predict an ordered layered ground state ( P 3 ¯ m 1 ) , structural characterization reveals a disordered cubic phase ( Fm 3 ¯ m ) stabilized by configurational entropy at synthetic temperatures. Furthermore, all-solid-state lithium batteries incorporating this electrolyte deliver a stable long-term cycling performance with a 90.6% capacity retention over 200 cycles at 0.5 C. This work demonstrates a closed-loop discovery paradigm where physics-constrained learning and thermodynamic understanding converge to realize high-performance, economically viable energy materials.