Small-sample ensemble learning-driven prediction and mechanistic analysis of hydrogen release performance in modified LiBH 4

Abstract

To address the complexity and high cost of traditional approaches for metal-modified LiBH4 systems, this study proposes a small-sample ensemble learning-density functional theory (EL-DFT) framework for the efficient prediction and mechanistic analysis of hydrogen dissociation energies ( E d ) in bimetal-doped structures (TM 1 TM 2( S i) Li 14 B 16 H 64 ). A training set was constructed incorporating structural geometric parameters as well as the thermophysical and chemical properties of the transition metals, from which 15 key features were selected through feature engineering. Gradient Boosting Regression (GBR) was employed as the base learner to establish the structure-performance relationship, and an AdaBoost framework was further introduced to construct an AdaBoost-GBR ensemble model to address the underfitting issue of the base learner. Model interpretability analysis indicated that the doping site ( Si ) of TM 2 is the key factors influencing E d . The ScSc ( S 1) Li 14 B 16 H 64 structure was screened out, and its favorable hydrogen release performance and thermal stability were confirmed via ab initio molecular dynamics (AIMD) simulations and projected density of states (PDOS) analysis. Meanwhile, the SISSO algorithm was employed to reveal the intrinsic relationships between local structures and performance. This study demonstrates that small-sample ensemble learning can effectively enhance model fitting and predictive performance under data-limited conditions.