EVENT DOSSIER
The researchers developed a new type of displacement-invariant Bayesian optimization method to address the issue of symmetry in the replacement between injection wells and production wells, which the standard Gaussian process kernel function cannot utilize in well location optimization for carbon capture and storage (CCS) projects. This method encodes displacement invariance by comparing the stability of the divergence between induced empirical representations and introduces a deep learning kernel learning model based on a depth set architecture as an invariant baseline. The method was evaluated in eight cases, including seven synthetic benchmarks and one real Johansen-structured CCS case, and its effectiveness was verified.
研究人员提出一种用于碳捕获与存储（CCS）项目井位优化的置换不变贝叶斯优化方法。该方法针对标准高斯过程核函数无法利用注入井和产出井组内置换对称性的问题，开发了新型高斯过程核（GP-Perm），通过比较诱导的经验表示之间的稳定发散度来编码置换不变性，并可与标准核结合处理向量输入；同时引入基于深度集架构的深度学习核学习模型（DKL-DS）作为学习到的不变基线。研究在八个用例中评估了该方法，包括七个合成基准和一个真实的 CCS 案例（Johansen 构造）。