Modeling heat capacity of liquid siloxanes using artificial intelligence methods

Abstract

In organic Rankine cycle (ORC) technology, heat capacity is an extremely important thermophysical property as it has significant implications for heat exchangers design as well as the determination of cycle performance. Given its importance, it is crucial that the heat capacity values used in the aforementioned applications are highly accurate, to ensure optimal performance and efficiency of ORC systems. Thus, the goal of this research is to develop precise data-driven models based on Random Forest, Decision Tree, AdaBoost, K-Nearest Neighbors, Ensemble Learning, Multilayer Perceptron Artificial Neural Network, Convolutional Neural Network and Support Vector Machine learning algorithms to accurately predict liquid siloxanes’ heat capacity as a function of molar mass, boiling point, temperature, and pressure. The findings suggest that nearly all the collected experimental data are technically appropriate for building the model. The analysis of sensitivity indicates a positive effect of temperature and negative effect of pressure, siloxane boiling point, and molar mass on the heat capacities. The evaluation metrics and graphical analyses indicate that the CNN-based model is the most accurate for the prediction task of heat capacity. It achieves the highest R-squared value (0.975042) and the lowest mean square error (0.000554) on the test set, along with the lowest average absolute relative error percentage (0.632605). These results highlight the CNN's superior ability to capture complex patterns and deliver precise predictions specifically for the prediction task of heat capacity.