2.5 Data set
The quality of a method depends largely on the quantity and quality of experimental data used in its development. The data for developing atomic connectivity group contribution method are from National Institute of Standards and Technology (NIST)32 , CRC Handbook of Chemistry and Physics33 and DIPPR80134, and only the experimental data are used for ACGC modelling work. We take the median value of repeated experimental test. And compounds with more than three carbon and silicon atoms combined were selected, because small molecules may affect the stability of the model, even though they can improve the accuracy of the model. If a group appears only in one or two compounds, the error for these compounds is small, while these groups are unreliable for predicting. In order to develop reliable model, the compounds with a group appearing in less than five compounds will be deleted. The data set is composed of 680 critical temperature (Tc ), 481 critical pressure (Pc ), 471 critical volume (Vc ).
Results and discussions
The results forTc ,P cand V c by ACGC methods are shown in Table 1. Statistical parameters such as square correlation coefficient (R 2), average absolute error (AAE), average relative error (ARE) were used to evaluate the ACGC model.R 2 is 0.9891, 0.9887, 0.9983 forTc ,P c and V c respectively. Leave-one-out cross-validation (LOO-CV) was applied for evaluating the predicting ability of ACGC models. The quality of ACGC method was then decided by the squared correlation coefficient of LOO-CV (Q 2). In AAG method, only the atomic adjacent group is used as traditional GC method. Compared with AAG and AAGS, theR 2 of ACGC model was improved. TheR 2 and ARE for the AAG models are shown in Table 1. The ARE of T c, P cand V c decreased by 42.57 %, 13.23 %and 6.82 %, respectively,when SF is applied. When ACF is applied to ACGC model, compared with AAGS model, it decreases by 34.27 %, 12.64 % and 34.63 %, respectively. Therefore, ACF is also effective in improving ACGC. ARE decreased by 24.19-62.25 % and AAE decreased by 19.30-61.54 % when SF and ACF are applied, therefore, SF and ACF are effective in improving ACGC method.
Table 1 The results for critical properties by ACGC method.