Scholarlink

<p>Stability modeling of halide perovskite solar cells (PSCs) is complicated by heterogeneous literature datasets and protocol-dependent variability. Here, a cross-model machine learning framework was developed to analyze PSC stability (<i>T</i><sub>80</sub>). Five regression models, including Random Forest (RF), Gradient Boosting, XGBoost, Support Vector Regression, and Linear Regression, were trained and interpreted using SHAP to prioritize mechanistic insight over predictive accuracy alone. Ensemble models achieved mean absolute errors of approximately 400–450&#xa0;h, reflecting intrinsic dataset variability. Domain-level attribution analysis showed that RF primarily emphasized compositional descriptors, whereas boosted models assigned greater importance to testing protocols and performance-related variables. Feature ranking remained stable after removal of publication-specific identifiers (Spearman <i>ρ</i> ≈ 0.83–0.90), supporting the robustness of intrinsic stability signals. Feature–feature interactions, quantified using the&#xa0;Pearson correlation of signed SHAP values, revealed strong composition–composition coupling (|<i>r</i>|&#xa0;&gt; 0.9) and pronounced halide–temperature interactions (<i>r</i> ≈ 0.87), while architecture–composition correlations remained weak. These findings demonstrate that explainable ML can differentiate genuine material-driven stability factors from proxy effects arising from heterogeneous experimental reporting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable machine learning analysis of stability mechanisms and feature interactions in halide perovskite solar cells

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Mohd Taukeer Khan,
  • Saleh O. Allehabi

摘要

Stability modeling of halide perovskite solar cells (PSCs) is complicated by heterogeneous literature datasets and protocol-dependent variability. Here, a cross-model machine learning framework was developed to analyze PSC stability (T80). Five regression models, including Random Forest (RF), Gradient Boosting, XGBoost, Support Vector Regression, and Linear Regression, were trained and interpreted using SHAP to prioritize mechanistic insight over predictive accuracy alone. Ensemble models achieved mean absolute errors of approximately 400–450 h, reflecting intrinsic dataset variability. Domain-level attribution analysis showed that RF primarily emphasized compositional descriptors, whereas boosted models assigned greater importance to testing protocols and performance-related variables. Feature ranking remained stable after removal of publication-specific identifiers (Spearman ρ ≈ 0.83–0.90), supporting the robustness of intrinsic stability signals. Feature–feature interactions, quantified using the Pearson correlation of signed SHAP values, revealed strong composition–composition coupling (|r| > 0.9) and pronounced halide–temperature interactions (r ≈ 0.87), while architecture–composition correlations remained weak. These findings demonstrate that explainable ML can differentiate genuine material-driven stability factors from proxy effects arising from heterogeneous experimental reporting.