A Closed-Loop Machine-Learning Framework for Inverse Design and Optimization of Carbon Capture and Storage and Underground Hydrogen Storage in Deep Saline Aquifers

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:A novel unified ML framework combining surrogate modeling, active learning, SHAP, and inverse design for subsurface applications.

🏢実務担当者:Enables efficient screening and optimization of CCS/UHS storage sites and operational parameters using ML-based surrogates.

🏛政策担当者:Demonstrates how advanced ML can reduce uncertainty and cost in CCS/UHS, supporting informed regulatory and investment decisions.

Existing surrogate models for carbon capture and storage (CCS) and underground hydrogen storage (UHS) are primarily designed for forward prediction and only rarely integrate temporal prediction, SHAP-based interpretability, and inverse design within a single framework. In this study, we propose a unified machine-learning (ML) framework that combines simulation, tabular panel regression, active learning, and evolutionary inverse design for CCS and UHS in deep saline aquifers. Four surrogate pipelines were developed for non-temporal and temporal CCS and UHS using 9,835, 9,999, 1,000, and 1,000 CMG GEM 2022.10 simulation cases, respectively. Best non-temporal performance was achieved by TabTransformerLite for CCS (nRMSE = 3.30 × 10⁻², R² = 0.920) and by CatBoost for UHS (R² = 0.994, nMSE = 5.68 × 10⁻³). In the temporal pipelines, FT-Transformer achieved the best overall performance for CCS (nRMSE(std)macro = 0.172; ~400× speedup relative to simulation), whereas NODE ranked first for UHS (nRMSE(std)macro = 0.058; ~428× speedup). Active learning reduced the weighted out-of-distribution slope by 34% for temporal CCS and lowered nRMSE(std)macro in temporal UHS by 14.0%, 10.3%, 5.2%, and 2.1% over four successive rounds relative to random sampling. SHAP analysis identified depth and permeability as the dominant CCS controls, whereas production duration and production rate were the main economic controls in UHS. Inverse design located a high-mineralization window for non-temporal CCS at a median depth of 2,252 m and yielded a median first-crossing time of 327.7 years for temporal CCS to reach the mineralization target. For non-temporal UHS, inverse design showed that each additional percentage point of hydrogen recovery was associated with an increase of about 9.1–9.6 × 10⁷ USD in upper-envelope NPV until diminishing returns emerged near 90% recovery. For temporal UHS, inverse design identified near-optimal solutions reaching NPV values of 3.5–4.0 × 10⁷ USD by the fifth storage cycle. Overall, the proposed framework advances subsurface ML from forward prediction toward decision-oriented support for CCS mineral-trapping enhancement and UHS economic optimization.

• openalex https://doi.org/10.11575/prism/51280first seen 2026-05-17 06:46:35 · last seen 2026-05-21 04:52:06