Predicting Heat Capacity of Molecular Fluids Using Interpretable Machine Learning Model

• DOI: 10.1021/acs.iecr.4c02495
• Publication Date: 2024-08-20T11:41:25Z
• Authors: Simin Li; Hongyan He; Yanlei Wang

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