Due to the extensive utilization of liquid nature gas (abbreviated as LNG) resources and a multitude of considerations, LNG storage tanks are gradually transitioning towards smaller footprints and heightened safety standards. Consequently, underground LNG storage tanks are being designed and constructed. However, underground LNG storage tanks release a considerable quantity of cold into the ground under both accidental and normal conditions. The influence of cold results in the ground freezing, which further compromises the safety of the structure. Existing research has neglected to consider the effects of this. This oversight could potentially lead to serious safety accidents. In this work, a complete set of experiments using a novel LNG underground storage tank fluid-solid-thermal coupled cryogenic leakage scale model were conducted for the first time to simulate the effect of the tank on the soil temperature field, stress field, and displacement field and to analyze the development of the three fields and the results of the effect. This research helps the related personnel to better design, construct, and evaluate the LNG underground storage tanks to avoid the catastrophic engineering risks associated with cryogenic leakage and helps to improve the design process of LNG underground storage tanks.

The thermal insulation integrity of liquefied natural gas storage tanks is essential for their life-cycle safety. However, perlite settlement (insulation material) can result in thermal leaks and lead to engineering risks. The direct measurement of perlite settlement is difficult due to the enclosed structure of these tanks. To address this challenge, this study presents a data-driven approach based on machine learning and real-time monitoring data. This study proposes a multi-fidelity machine learning framework to enhance generalizability and leverage multi-fidelity data effectively. Low-fidelity data are readily available but contain systematic errors, while high-fidelity data are accurate but limited in accessibility. By combining both types of data, this framework enhances the generalisability and prediction accuracy of trained models. The results of the data experiments demonstrate that the multi-fidelity framework outperforms models trained solely on low- or high-fidelity data, achieving a coefficient of determination of 0.980 and a root mean square error of 0.078 m. Three machine learning algorithms—Multilayer Perceptron, Random Forest, and Extreme Gradient Boosting—were evaluated to determine the optimal implementation. This approach provides a reliable method for the real-time monitoring of thermal insulation integrity in liquefied natural gas storage tanks, contributing to improved industrial safety and operational efficiency.