Volume 55

Utilizing Transfer Learning for Enhanced Prediction of Supercritical CO2 Critical Flows Haifan Liao, Zhicheng Liang, Hongfei Hu, Haijun Wang

Abstract

With the widespread adoption of Carbon Capture and Storage (CCS) technologies and investigations into advanced power cycles such as the supercritical CO2 (sCO2) Brayton cycle, it is particularly crucial to focus on the risks and dynamics associated with accidental leaks under supercritical conditions. One key aspect is the accurate prediction of critical flow rates of supercritical fluids during leakage events. Owing to the scarcity of experimental data on critical flow in sCO2 microchannels, this study harnesses the abundant experimental data from critical flow studies of water, employing machine learning algorithms enhanced by transfer learning to predict the critical flow rates of sCO2. The approach involves pre-training a neural network on critical flow data from water, followed by fine-tuning with sCO2 data, thereby bridging the gap in data availability and enhancing the model’s generality across various parameters. The principal research findings indicate that transfer learning can improve prediction accuracy and adaptability, suggesting that transferring knowledge from extensively studied fluids like water can effectively enhance the predictive performance of less-documented supercritical fluids. This research provides a valuable tool for designing safer CCS and energy systems.

Keywords Carbon Capture and Storage; Supercritical Carbon Dioxide; Critical Flow; Neural Networks; Transfer Learning.