CIESC Journal ›› 2024, Vol. 75 ›› Issue (2): 429-438.DOI: 10.11949/0438-1157.20230955

• Thermodynamics • Previous Articles     Next Articles

Prediction of self-diffusion coefficients of ionic liquids using back-propagation neural networks

Yongjun XIAO(), Zhaochong SHI, Ren WAN, Fan SONG, Changjun PENG(), Honglai LIU   

  1. School of Chemistry and Molecular Engineering, East China University of Science and Technology, Shanghai 200237, China
  • Received:2023-09-13 Revised:2024-01-09 Online:2024-04-10 Published:2024-02-25
  • Contact: Changjun PENG

反向传播神经网络用于预测离子液体的自扩散系数

肖拥君(), 时兆翀, 万仁, 宋璠, 彭昌军(), 刘洪来   

  1. 华东理工大学化学与分子工程学院,上海 200237
  • 通讯作者: 彭昌军
  • 作者简介:肖拥君(1998—),女,硕士研究生,xiaoyongjun00@163.com
  • 基金资助:
    国家自然科学基金项目(22078086)

Abstract:

Using the charge density distribution fragment area (Sσ ) and hole volume (VCOSMO) obtained by the fragment activity coefficient conductor-like shielding model (COSMO-SAC) as structural descriptors, we developed a quantitative structure-property relationship (QSPR) model, namely the BP-ANN model, to predict cation and anion self-diffusion coefficients of ionic liquids. The range of applicability and predictive capability of the BP-ANN model were also examined and compared with another QSPR model established by linear regression (Model I). The results revealed that the BP-ANN model can be applied to a broader range of ionic liquid species compared with Model I. The BP-ANN model achieves a high coefficient of determination (R2) value exceeding 0.99 in the training, validation, and testing dataset for cations, and surpassing 0.98 for anions across all sub-datasets. For the total dataset, the BP-ANN model yields low average absolute relative deviations (AARD) of 2.8% for cations and 3.7% for anions between calculated and experimental values, while the corresponding values for Model I are 14.54% and 14.57%, respectively. Therefore, the prediction performance of the BP-ANN model is significantly better than that of the model based on linear regression.

Key words: neural networks, QSPR, ionic liquids, self-diffusion coefficient, prediction

摘要:

以片段活度系数类导体屏蔽模型(COSMO-SAC)获得的电荷密度分布片段面积(Sσ )和空穴体积(VCOSMO)为结构描述符,采用反向传播神经网络构建了可用于预测离子液体中阴阳离子自扩散系数的定量结构-性质关系(QSPR)模型——BP-ANN模型。考察了模型的适用范围与预测能力,并与线性回归得到的QSPR模型(模型I)进行了比较。发现BP-ANN模型适用的离子液体种类较广,模型在阳离子自扩散系数训练集、验证集与测试集中的决定系数R2均大于0.99,在阴离子的三个子集中的R2均大于0.98。阳离子与阴离子预测的平均绝对相对误差分别为2.8%和3.7%。线性回归的QSPR模型对应的值分别为14.54%和14.57%,即BP-ANN模型的预测效果要明显优于基于线性回归建立的模型。

关键词: 神经网络, 定量构效关系, 离子液体, 自扩散系数, 预测

CLC Number: