Abstract As geological CO2 storage moves from pilot demonstrations toward large-scale deployment, storage projects require faster, more transparent and auditable decisions under strong geological heterogeneity and operational uncertainty. Machine learning is increasingly used as a set of workflow accelerators that integrate data, simulation, monitoring and engineering decisions. This review evaluates machine learning across four lifecycle stages: site screening, injection design, operational monitoring and monitoring, reporting and verification accounting. Instead of organising the literature by algorithm family, it uses engineering tasks, performance metrics and deployable workflows as the main framework. The review summarises stage-specific data streams, learning tasks, validation needs and outputs, and identifies three workflow families: surrogate-enabled scenario evaluation and optimisation, assimilation-oriented hybrid state updating, and integrated workflows for monitoring, reporting and verification. It also compares machine-learning and physics-based approaches and outlines requirements for trustworthy deployment, including cross-scenario validation, explicit uncertainty quantification, applicability-domain control, explainability and workflow governance. The analysis shows that the most useful role of machine learning is not to replace numerical simulation, but to shorten engineering iteration loops while retaining physical consistency, uncertainty transparency and auditable evidence. Future progress will depend on open benchmarks, multimodal data fusion, hybrid physics-machine learning workflows and governance structures that can support regulatory review and long-term stewardship.
Wang et al. (2026) studied this question.