Carbon capture, enhanced oil recovery (EOR)-utilization and storage (CCUS-EOR) is recognized as an effective approach to mitigate greenhouse gas emissions while delivering economic benefits. However, its practical deployment is limited by the absence of advanced deep learning models for petroleum tabular data, the limited adaptability of existing optimization methods, and the lack of comprehensive evaluation for full-cycle CCUS-EOR. Here, we introduce a generalizable framework that integrates mechanism experiments, numerical simulations, and deep learning methods to address these challenges. Three-stage experiments are conducted to clarify microscopic displacement mechanisms and provide key parameters for numerical simulation. Based on field-scale simulations of 20 years of CO2 water-alternating-gas (WAG) injection followed by 19 years of pure CO2 storage until 2060, we develop a TabPFN-based meta-learning surrogate model for joint prediction of oil recovery, CO2 storage, and net present value (NPV), achieving high accuracy (prediction error <2%, R2 > 0.97) compared to baseline models. We further apply an improved multi-objective optimization using the Adaptive Crossover and Adaptive Mutation Non-dominated Sorting Genetic Algorithm II (ACAM-NSGA-II) to obtain optimal Pareto solutions. Compared to baseline cases, the proposed framework significantly enhances CCUS-EOR performance, enhancing oil recovery by 27.05% (from 5.95 × 105 t, 35.17% to 1.05 × 106 t, 62.22%), tripling CO2 storage capacity (from 1.33 × 106 to 4.45 × 106 t), and improving NPV by 68.0% (from $344 million to $578 million). The Pareto front is further divided into three different solution regions, thereby elucidating the underlying physical mechanisms associated with each cluster and providing clear operational insights for target-oriented CO2-WAG design. This study offers a scalable blueprint framework for large-scale engineering design in petroleum engineering, particularly in tabular prediction and multi-objective optimization contexts.