Develop machine learning-based model and automated process for predicting liquid heat capacity of organics at different temperatures

Heat capacity of compounds is an important reference parameter in thermodynamic, chemical processes, energy storage, heat transfer, heat exchange. However, the heat capacity data of most liquid compounds are missing due to the limited availability of data and the complexity of its experimental measurement. Heat capacity prediction models established did not consider temperature influences, or their complex prediction processes restricted their application. In this work, we developed an automated process based on machine learning algorithms to directly predict the liquid heat capacity of compounds from their molecular structure, with broader applicability to different temperature conditions. A dataset including 1253 heat capacity samples of liquid compounds was collected and divided into training set and test set in a 7:3 ratio. Random Forest (RF), Gradient Boosting Regression Tree (GBRT), Multi-Layer Perception (MLP), and Ridge Regression (RR) algorithms were utilized to develop predictive models. Ten-fold cross-validation, combined with the coefficient of determination (R2), was employed to evaluate the model performance. Notably, GBRT performed the best on the test set with an R2 of 0.973, an AARD% of 5.248, and an MAE of 9.385 J/mol•K demonstrates the high accuracy of the model, and the model was saved and integrated into an automated prediction process. Meanwhile, through the application of the SHAP method, it was discovered that ATS0pe, NumValenceElectrons, and Kappa2 were identified as the most important features. This study introduces an automated approach for predicting the liquid heat capacities of compounds at various temperatures for the researchers in related fields.