基于数值模拟与机器学习融合的CO₂封存研究

Research on CO₂ Sequestration Based on the Integration of Numerical Simulation and Machine Learning

摘要: 随着全球碳排放持续增长加剧气候变化，二氧化碳地质封存(CCUS)技术的重要性日益凸显，其中深部咸水层和煤层因其巨大潜力成为主要封存场所。融合数值模拟与机器学习技术成为突破瓶颈的关键。本文综述了该融合方法在碳封存领域的研究进展与应用。数值模拟方面，以COMSOL为代表的多物理场耦合软件能够精准模拟CO₂在多孔介质中的运移如两相流、溶解扩散、岩石变形等。煤层吸附膨胀效应，其数学模型构建、关键参数设定与边界条件处理是可靠模拟的基础。机器学习方面，支持向量机、随机森林和神经网络等算法，利用数值模拟生成的数据集，可高效构建算法模型，显著加速参数优化、不确定性分析、封存量预测及稳定性评估。工程实例验证了该融合框架的显著优势，通过代理模型将历史拟合计算量降低数个数量级。研究表明，数值模拟与机器学习的协同大幅提升了碳封存预测的效率和精度，解决了强非均质性和超长期预测难题，为CCUS工程的高效决策与优化提供了强有力的技术支撑，有力推动了碳中和目标下油气开发与碳封存技术的融合发展。

Abstract: As global carbon emissions continue to increase and climate change intensifies, the importance of carbon dioxide geological storage (CCUS) technology is becoming increasingly prominent, with deep saline and coal seams becoming the main storage sites due to their enormous potential. Integrating numerical simulation and machine learning technology has become the key to breaking through bottlenecks. This article reviews the research progress and application of this fusion method in the field of carbon sequestration. In terms of numerical simulation, multi physics coupling software represented by COMSOL can accurately simulate the transport of CO₂ in porous media, such as two-phase flow, dissolution diffusion, rock deformation, etc. The mathematical model construction, key parameter setting, and boundary condition treatment of coal seam adsorption expansion effect are the basis for reliable simulation. In terms of machine learning, algorithms such as support vector machines, random forests, and neural networks can efficiently construct algorithm models using datasets generated from numerical simulations, significantly accelerating parameter optimization, uncertainty analysis, storage prediction, and stability evaluation. Engineering examples have verified the significant advantages of this fusion framework, reducing the computational complexity of historical fitting by several orders of magnitude through proxy models. Research has shown that the synergy between numerical simulation and machine learning has significantly improved the efficiency and accuracy of carbon sequestration prediction, solved the problems of strong heterogeneity and ultra long term prediction, provided strong technical support for efficient decision-making and optimization of CCUS engineering, and effectively promoted the integrated development of oil and gas development and carbon sequestration technology under the goal of carbon neutrality.