CDS&E: Sustained risk mitigation of carbon storage with seismic monitoring through simulation and Bayesian inference

气候小组一致认为，碳捕获和封存(CCS)是减少大气中温室气体和应对气候变化的任何综合计划的重要组成部分。地下水库的二氧化碳泄漏风险被用于储存，阻碍了对CCS的更广泛接受。为了帮助管理风险，任何CCS努力都必须使用监测系统来检测泄漏，例如地震监测。该项目解决了地震监测系统开发中的几个挑战，该系统可持续到不确定的未来。这些挑战包括来自永久性和低成本传感器的噪声数据，描述地下二氧化碳羽流所需的高维数据，需要对不确定性进行量化以向利益攸关方保证，以及连续监测系统的泄漏自动检测。这个项目包括扩大对计算科学的参与的活动，包括开发和部署一个面向不同背景、兴趣和职业目标的学生的科学应用机器学习短期课程。这个项目的技术目标是开发一种结合延时地震成像、物理模拟和不确定性量化的方法来监测地下蓄水池中的临界二氧化碳羽流。这项工作通过使用贝叶斯数据同化和二氧化碳羽流的流体流动模拟来扩展当前的实践。为了便于处理，使用条件可逆神经网络来表示复杂的概率分布。数据同化还与联合恢复方法相结合。二氧化碳羽流的后验分布被输入到生成式分类器中，以自动检测泄漏，同时估计不确定性。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Carbon capture and storage (CCS) is agreed upon by climate panels to be an essential part of any comprehensive plan for reducing greenhouse gases in the atmosphere and combating climate change. Broader acceptance of CCS is hindered by the perceived risk of carbon dioxide leakage from the underground reservoir being used for storage. To help manage risk, any CCS endeavor must use a system of monitoring to detect leaks, such as seismic monitoring. This project addresses several challenges in the development of a seismic monitoring system that is sustainable into the indeterminate future. These challenges include noisy data from permanent and low-cost sensors, high-dimensional data needed to describe the carbon dioxide plume underground, the need for uncertainty quantification to assure stakeholders, and automated detection of leaks for continuous monitoring systems. This project includes activities to broaden participation in computational science, including the development and deployment of a short course on machine learning for scientific applications geared to students with diverse backgrounds, interests, and career goals.The technical aim of this project is to develop a methodology for monitoring critical carbon dioxide plumes in underground storage reservoirs that combines time-lapse seismic imaging, physical simulation, and uncertainty quantification. The work extends current practice by using Bayesian data assimilation with a fluid flow simulation for the carbon dioxide plume. For this to be tractable, conditional invertible neural networks are used to represent complex probability distributions. The data assimilation is also coupled with joint recovery methods. The posterior distribution of the carbon dioxide plume is input into a generative classifier to automatically detect leakages while also estimating uncertainty.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.