Scholarlink

<p>Accurate prediction of thermodynamic properties such as enthalpy of formation (ΔH<sub>f</sub>) is essential for designing chemical compounds with targeted stability and performance. In this study, a computational screening of polycyclic aromatic hydrocarbons (PAHs) was conducted using machine learning (ML) models trained on experimentally reported ΔH<sub>f</sub> values. Over 200 molecular descriptors were calculated via RDKit, with feature selection guided by correlation analysis to retain the most relevant parameters. Twelve ML algorithms were evaluated, with CatBoost and XGBoost demonstrating superior predictive accuracy, achieving R² scores above 0.95 on validation sets. These optimized models were applied to the COMPAS database of ~ 34,000 PAHs to identify candidates with low predicted ΔH<sub>f</sub> values. Fifty top-ranked compounds were shortlisted based on predicted thermodynamic stability, synthetic accessibility (SA), and chemical similarity network (CSN) analysis. Selected candidates exhibited SA scores between 1.4 and 1.9, suggesting feasible synthesis routes. Visualization via t-SNE and CSN highlighted structural diversity while confirming clustering of low-ΔH<sub>f</sub> compounds. The approach demonstrates the integration of ML prediction, synthetic feasibility assessment, and chemical space mapping as an efficient strategy for targeted molecular discovery.</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

Exploration and Screening of Database of Polycyclic Aromatic Hydrocarbons to Find Compounds with Lower Enthalpy of Formation

  • 期刊论文

  • 发表日期：

    2026年7月18日

查看全文
  • Maymounah N. Alharthi,
  • Tagir Kadyrov,
  • Khadijah Mohammedsaleh Katubi,
  • M. S. Al-Buriahi

摘要

Accurate prediction of thermodynamic properties such as enthalpy of formation (ΔHf) is essential for designing chemical compounds with targeted stability and performance. In this study, a computational screening of polycyclic aromatic hydrocarbons (PAHs) was conducted using machine learning (ML) models trained on experimentally reported ΔHf values. Over 200 molecular descriptors were calculated via RDKit, with feature selection guided by correlation analysis to retain the most relevant parameters. Twelve ML algorithms were evaluated, with CatBoost and XGBoost demonstrating superior predictive accuracy, achieving R² scores above 0.95 on validation sets. These optimized models were applied to the COMPAS database of ~ 34,000 PAHs to identify candidates with low predicted ΔHf values. Fifty top-ranked compounds were shortlisted based on predicted thermodynamic stability, synthetic accessibility (SA), and chemical similarity network (CSN) analysis. Selected candidates exhibited SA scores between 1.4 and 1.9, suggesting feasible synthesis routes. Visualization via t-SNE and CSN highlighted structural diversity while confirming clustering of low-ΔHf compounds. The approach demonstrates the integration of ML prediction, synthetic feasibility assessment, and chemical space mapping as an efficient strategy for targeted molecular discovery.