Using machine learning in QSPR to estimate the boiling and critical temperatures of pure organic compounds

Estimating physical and chemical properties, such as boiling temperature (Tb) and critical temperature (Tc), for organic compounds remains a significant challenge in chemical engineering. Accurate prediction of these properties has been a major research focus due to their importance in various applications. This study aims to develop models using a Quantitative Structure-Property Relationship (QSPR) approach to predict Tb and Tc for 417 and 412 organic compounds, respectively. The models rely on a machine learning algorithm, the multi-layer perceptron artificial neural network (MLP-ANN), for nonlinear modeling based on relevant molecular descriptors as input variables. A comparison with support vector regression (SVR) was conducted to assess the effectiveness of MLP-ANN. The optimal configurations for the MLP-ANN models were (25-17-1) for Tb and (25-14-1) for Tc. Various statistical metrics, R², IOA, MAE, MAPE, and RMSE, were used to measure model accuracy and stability. For the MLP-ANN Tb model, results included R² = 0.9974, IOA = 0.9992, MAE = 3.6331, MAPE = 1.0165, and RMSE = 4.9321. For the Tc model, results were R² = 0.9935, IOA = 0.9982, MAE = 7.0545, MAPE = 1.0436, and RMSE = 9.5482. The MLP-ANN models consistently outperformed the SVR models, demonstrating superior accuracy, stability, and generalization. Additionally, the applicability domain (AD) analysis confirmed the reliability and generalizability of the models, with most data points falling within an acceptable range. A comparison with previous models showed that the proposed models surpass them in precision and robustness, highlighting the strong capability of models MLP-ANN to provide accurate predictions.