Interpretable Gradient Boosting for Vanadium Redox Flow Battery Energy Efficiency: Leakage-Aware Validation and Comparison with Gaussian Processes

Vanadium redox flow battery (VRFB) energy efficiency is affected by interacting cell-design and operating variables. We re-evaluated a published 420-record, 12-descriptor dataset with gradient boosting decision trees (GBDTs), audited the original hyperparameter-search implementations, and performed repeated and setting-group holdout validation. The original 90/10 partition comprised 378 development and 42 test records. Four published parameter configurations gave single-split test R² values of 0.8870, 0.8459, 0.8726, and 0.8124, respectively. Code audit identified Halton quasi-Monte Carlo sampling in the originally QMC-labelled run and CMA-ES sampling in the other three runs; therefore the original four-way comparison cannot support claims about four distinct Bayesian/evolutionary optimization algorithms. In three repetitions of five-fold cross-validation with fixed hyperparameters, the QMC-associated configuration achieved mean R² = 0.8401 (SD 0.0298); 12 repeated random holdouts gave 0.8188 (SD 0.0823). When complete operating-setting groups (excluding current density) were kept together, its mean R² decreased to 0.4555 (SD 0.3059; 10 holdouts). A same-split Matérn-5/2 Gaussian process baseline obtained test R² = 0.9252, while the published GPR R² of 0.9565 used a different evaluation protocol and is not directly comparable. SHAP rankings for two GBDT configurations were strongly concordant (Spearman ρ = 0.972), with current density and electrode treatment ranking first and second. The findings support interpretable, exploratory screening but show that stronger provenance-aware validation is required before claiming transferable deployment performance.

Hali tarjima qilinmagan