Predictive analysis of CO₂ hydrate phase equilibrium temperature based on CKAN learning algorithm and Shapley additive explanations

Abstract In response to the increasingly severe issue of anthropogenic carbon dioxide (CO 2 ) emissions, CO 2 hydrate storage presents a promising method for reducing atmospheric pollution and mitigating climate change. This study proposes a machine learning model based on the Convolutional-Kolmogorov-Arnold (CKAN) neural network, aiming to accurately predict the phase equilibrium temperatures of CO 2 hydrates in systems containing salts, various promoters, and inhibitors. This study utilized 1909 experimentally measured equilibrium data points of hydrate temperatures, covering systems with salts and thermodynamic agents. After preprocessing, 1527 data points were selected for model training, and 382 data points were used for model validation. The CKAN model involves 29 input variables and 1 output variable. Its predictive accuracy was evaluated using various statistical error metrics, and its effectiveness and interpretability were examined using methods such as Shapley additive explanations (SHAP) to identify features with significant positive or negative impacts on the model’s predictions. Compared to the traditional convolutional neural (CNN) network and multi-layer perceptron (MLP) models, long short-term memory (LSTM) networks, light gradient boosting machine (LightGBM) model, and natural gas hydrate (NGH) phase equilibrium calculation software, the CKAN model exhibits a significant accuracy advantage. The CKAN model can serve as a reliable tool for predicting the phase equilibrium curves of CO 2 hydrates in pure water, single-salt brine, multi-salt brine, and under various conditions.