Heat Capacity of Ionic Liquids: Toward Interpretable Chemical Structure-Based Machine Learning Approaches

This
study focuses on predicting the heat capacity of pure liquid-phase
ionic liquids (ILs) using machine learning models from various categories,
including support vector machines, instance-based learning, ensemble
learning, and neural networks, with linear regression serving as a
baseline. A key aim of this work is not only to achieve accurate predictions
but also to ensure the interpretability of the results, addressing
a gap often overlooked in predictive modeling studies. To accomplish
this, we curated and cleaned a comprehensive data set of 13,893 data
points covering 322 ILs, using temperature and chemical structure-based
features as inputs. We evaluated model performance and conducted a
thorough interpretability analysis to reveal the patterns of the top-performing
model’s predictions, ensuring that they are understandable.
All models outperformed the baseline, with XGBoost (eXtreme Gradient
Boosting) from the ensemble learning category achieving the best results,
with total RMSE, R2, and AARD (%) values
of 11.389, 0.997, and 1.212%, respectively. Shallow neural networks
also performed competitively, suggesting that complex deep learning
architectures may not be necessary. Both 10-fold and leave-one-IL-out
(LOILO) cross-validation further validated the robustness of these
results. Importantly, the interpretability analysis identified key
factors influencing heat capacity predictions, such as anion size
(e.g., NTf2 and FAP) and alkyl chain length. These factors
were validated by testing the model on previously unseen IL examples.
Additionally, a user-friendly web application was developed to make
predictions, allowing users to input chemical groups or select compounds
from a predefined list of 1633 ILs. This study underscores the importance
of combining diverse modeling approaches with robust interpretability
techniques to achieve reliable and explainable predictions for IL
heat capacity.