Feature analysis of generic <scp>AI</scp> models for <scp>CO<sub>2</sub></scp> equilibrium solubility into amines systems
Ting Lan, Shoulong Dong, Hui Luo, Liju Bai, Helei Liu
Abstract
Abstract Reported models have disadvantages such as poor prediction accuracy and time‐consuming. And they can not reflect the impact of chemical reactions on CO 2 solubility. To compensate for these deficiencies, parameters representing operational parameters, physical properties, chemical properties, and molecular properties are introduced as input variables. A series of models are constructed by three algorithms: back propagation neural network, radial basis function neural network, and random forest. The model with the best prediction performance is level OPCM (RBFNN), with the AARE of only 1.52%. By ranking the importance of the features using the RF algorithm, P CO2 , was found to be the key parameter affecting the CO 2 loadings, with M being the least important. Using the screened key parameters to model the model, as well as optimizing the structure, can further improve the predictive performance of the model. The full process development and optimization model framework constructed in this article can provide practical guidance for the development of machine learning models.