here, for data screening method: “no” refers to the model developed by
the initial dataset obtained from NIST; “post” means that the dataset
used to build the model is consistent with “no”, but only the data
points selected under the pre-screening rules are used for calculatingQ 2. “pre” refers the model with the dataset
after data pre-screening.
3.3.3. Comparison with references
The comparison of ρ (T ,P ,I )-QSPR model with
that in references was mainly carried out through,R 2 and Q 2, and the
details are shown in Table 4. Our model has advantages in data set
screening, rather than collecting too many data points. Moreover, theρ (T ,P ,I )-model has higher accuracy withR 2 of 0.9922. The R 2is slightly lower than that of our previous studies (Yan et
al.37 and Zhang et. al29), which can
be attributed to that the data pre-screening process in this work
avoided the excessive proportion of data points from the same ILs, so as
to effectively avoid the situation of high R 2.
In addition, the model was validated by LOCO-CV and LOAO-CV, which
proved the stability and robustness of theρ (T ,P ,I )-model. To sum up, this model is
reliable in calculating the ρ values for ILs.
Table 4 . Comparisons of this work with references for the
density.