Predicting Heat Capacity of Molecular Fluids Using Interpretable Machine Learning Model

Heat capacity at constant pressure (Cp) of a molecular liquid medium is not only a basic physical property applicable in the calculation of microscopic characteristics but also a crucial property in chemical engineering processes. In this work, we utilized machine learning (ML) methodologies based on an established molecular liquid database to develop predictive models for Cp of fluids. The training data for these models were sourced from the literature and densityFfunctional theory (DFT) calculations, with simplified molecular input line entry system (SMILES) strings encoded via one-hot encoding as inputs to our models. Interestingly, the traditional ML algorithm extreme gradient boosting (XGBoost) produced the most effective predictive model, achieving a Pearson correlation coefficient exceeding 0.99, a root mean squared error (RMSE) of 1.95, and a mean absolute error (MAE) of 0.89, thus providing more accurate predictions of the Cp than those obtained from DFT calculations. Additionally, we integrated Deep Belief Networks (DBNs) with Deep Neural Networks (DNNs), using DBNs for feature extraction and DNNs for subsequent prediction tasks. This combination successfully applied deep learning algorithms to small data set predictions, achieving a Pearson correlation coefficient of 0.96, a RMSE of 5.01, and a MAE of 2.21. Furthermore, we employed the SHAP library to analyze the interpretability of our models, examining the contribution of each feature within each descriptor to the overall model. Finally, comprehensive regression diagnostics were used to assess the reliability of the model, enhancing our understanding of its predictive capabilities.