Transfer Learning-Enabled Discovery of MXenes as Efficient Anchoring Materials for Li-S Batteries

Abstract

The practical deployment of lithium-sulfur (Li-S) batteries is severely constrained by the polysulfide shuttle effect, which leads to rapid capacity fading and poor cycling stability. Developing efficient host materials to strongly anchor soluble lithium polysulfides is thus crucial for performance enhancement. Here, a novel cross-property transfer learning framework is established to accelerate the discovery of two-dimensional transition-metal carbides/nitrides (MXenes) as highperformance anchoring materials. Using density functional theory, a source domain of 20,738 substrate energy samples and a target domain of 329 Li₂S₆ adsorption energy samples are constructed. A robust atomic-feature-augmented crystal graph convolutional neural network (AC-CGCNN) incorporating a channel attention mechanism is pre-trained on substrate energies and fine-tuned on adsorption energies, achieving excellent accuracy (R² = 0.97, MAE = 0.27 eV). Screening 556 MXenes identifies 115 candidates with optimal adsorption strengths (-1.3 to -2.7 eV). Among them, ScWCCl₂ and YNbNBr₂ exhibit exceptional dynamic stability and low rate-determining-step energy barriers, highlighting their potential as efficient LiPS anchors. This study demonstrates the effectiveness of transfer learning for accurate adsorption energy prediction under data-limited conditions and establishes a generalizable paradigm for accelerating the discovery of high-performance host materials for Li-S batteries.