基于ANN-QSPR算法的新型纯碳水化合物燃料性质预测方法

ANN-QSPR models for the predication of physical properties of a new-type carbohydrate fuel

开发新型纯碳水化合物燃料作为新能源时,必须预测和筛选纯碳水化合物的物理性质,从而找到可能合适的化学物质,然而由实验来逐一确定大量分子的物理性质既耗时又昂贵。研究发现,运用人工智能网络—定量构效关系(ANN-QSPR)算法来建立纯碳水化合物物理性质的计算模型可以起到事半功倍的效果。基于DIPPR 801数据库中的纯组分性质和DragonX软件包计算了相应碳水化合物的分子描述符,所构建的模型结合了定量构效关系(QSPR)和两层前馈人工智能网络(ANN)。由此建立了多个全面而可靠的模型来预测新型纯碳水化合物燃料的各种物理性质,包括正常沸点、闪点、燃烧焓、蒸发焓、液体密度、表面张力、液体的黏度和熔点等。为了提高模型中数据集之间的一致性,还引入了主成分分析法(PCA),以进一步消除分子描述符值的维数。另外,通过共识建模进行交叉验证,减少了不确定性的影响,提高了模型的预测精度。

When a pure carbohydrate compound is developed as a novel combustion fuel, its physical properties of each component will be first necessarily predicted and screened, However, the experimental determination of these properties for a huge amount of molecules can be very time consuming and costly. In view of this, the artificial neural network - quantitative structure-property rela- tionships （ANN-QSPR） algorithm was applied to build the desired models. Molecular descriptors were calculated based on a large number of pure components with evaluated values in DIPPR 801 database and the software package DragonX. The models developed were combinations of QSPR and two layer feed forward ANN. Thus the relatively comprehensive and reliable models were developed for predicting physical properties, including normal boiling point, flash point, enthalpy of combustion, enthalpy of vaporization, liq- uid density, surface tension, liquid viscosity, melting point, etc. For improving the consistency, principal component analysis （PCA） was introduced to further eliminate the dimensions of molecular descriptor values. Finally, the idea of cross validation for consensus modeling is further utilized to improve the predictive quality of obtained models.