Scholarlink

<p>Infinite dilution activity coefficient (<i>γ</i><sup>∞</sup>) is a key thermodynamic parameter in solvent design for chemical processes. Although conductor-like screening model for segment activity coefficient (COSMO-SAC) exhibits strong prior predictive capabilities, its estimations are sometimes only qualitative rather than quantitative. Another limitation of COSMO-SAC arises from the reliance on time-intensive quantum chemistry calculations, which restricts its scalability for large-scale solvent screening. To overcome these issues, this study integrates COSMO-SAC with machine learning for accurate <i>γ</i><sup>∞</sup> prediction of binary mixtures. By bypassing the necessity for quantum chemistry calculations, the multi-task machine learning model could rapidly predict the surface charge density distribution (<i>σ</i>-profiles) and molecular cavity volume (<i>V</i><sub>COSMO</sub>) of molecules and ions, while accurately distinguishing isomers. Four adjustable parameters of COSMO-SAC are optimized using more than 20000 experimental data points of <i>γ</i><sup>∞</sup>, and residual systematic errors are further corrected with the boosting ensemble strategy to improve the model performance. The resulting hybrid model reduces the mean absolute error from 0.944 to 0.102 (<i>R</i><sup>2</sup> = 0.969), representing an 89 % improvement, while preserving the physicochemical interpretability of model. This accurate and efficient approach broadens the practical applicability of <i>σ</i>-profiles and <i>V</i><sub>COSMO</sub> prediction, as well as <i>γ</i><sup>∞</sup> calculations based on COSMO-SAC, facilitating the high-throughput solvent screening for diverse chemical engineering applications.</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

Bridging machine learning and COSMO-SAC for accurate prediction of infinite dilute activity coefficients of binary mixtures

  • 期刊论文

  • 发表日期：

    2026年1月1日

查看全文
  • Yuxin Qiu,
  • Guzhong Chen,
  • Qian Liu,
  • Zhiwen Qi,
  • Kake Zhu,
  • Zhen Song

摘要

Infinite dilution activity coefficient (γ∞) is a key thermodynamic parameter in solvent design for chemical processes. Although conductor-like screening model for segment activity coefficient (COSMO-SAC) exhibits strong prior predictive capabilities, its estimations are sometimes only qualitative rather than quantitative. Another limitation of COSMO-SAC arises from the reliance on time-intensive quantum chemistry calculations, which restricts its scalability for large-scale solvent screening. To overcome these issues, this study integrates COSMO-SAC with machine learning for accurate γ∞ prediction of binary mixtures. By bypassing the necessity for quantum chemistry calculations, the multi-task machine learning model could rapidly predict the surface charge density distribution (σ-profiles) and molecular cavity volume (VCOSMO) of molecules and ions, while accurately distinguishing isomers. Four adjustable parameters of COSMO-SAC are optimized using more than 20000 experimental data points of γ∞, and residual systematic errors are further corrected with the boosting ensemble strategy to improve the model performance. The resulting hybrid model reduces the mean absolute error from 0.944 to 0.102 (R2 = 0.969), representing an 89 % improvement, while preserving the physicochemical interpretability of model. This accurate and efficient approach broadens the practical applicability of σ-profiles and VCOSMO prediction, as well as γ∞ calculations based on COSMO-SAC, facilitating the high-throughput solvent screening for diverse chemical engineering applications.