4. Conclusions
Two f (T ,P ,I )-QSPR models based on 19335ρ data points and 9238 η data points were established to
calculate the ρ and η of ILs under variable temperature
and pressure. In order to accurately verify the stability as well as the
robustness of the f (T ,P ,I )-QSPR model, a new
internal validation method, LOIO-CV, was proposed, which has more strict
evaluation criteria and is more reliable compared with the traditional
LOO-CV. Moreover, the stability of thef (T ,P ,I )-QSPR model is improved by using
data pre-screening to equalize the distribution of data points of ILs
and introducing descriptors to the temperature and pressure terms.
Furthermore, using only atomic linkage relationships of ILs, no
additional time is required to obtain the optimal structure of the ILs.
The results of statistical parameter analysis show that these models
have good prediction accuracy and reliability with highR 2 and low MAE. In the evaluation of ILsf (T ,P ,I )-QSPR model, both models passed the
rigorous LOIO-CV, which indicates that thesef (T ,P ,I )-QSPR models have good stability,
robustness, and relatively accurate prediction performance. Meanwhile,
we also found that the evaluation results of LOIO-CV were affected by
the distribution of ion species, that is, the model established by the
dataset with a more balanced distribution of data points of different
ILs had higher stability and robustness. Therefore, the equilibrium
distribution of different ILs data points is particularly important when
modeling the properties of ILs by using QSPR method. In one sense, two
proposed f (T ,P ,I )-QSPR models are widely
applicable to predict the ρ and η properties of ILs, and
these models provide an intelligent tool for predicting the design or
synthesis of ILs containing novel cations and anions. It is worth
mentioning that the strategy can be widely applied to the estimation of
other properties of ILs, such as environmental toxicity and other
related physicochemical properties.