Artificial intelligence-driven forecast of methane recovery and CO2 storage efficiency for carbon-neutral energy production from marine gas hydrates

This study develops an artificial-intelligence (AI) framework to predict methane recovery and carbon-dioxide (CO 2 ) storage efficiency during CH 4 /CO 2 replacement in marine hydrate–bearing sediments. Six models—Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), a hybrid CNN–LSTM, and ensemble learners (Random Forest, Gradient Boosting, Extra Trees) were trained on 277 peer-reviewed laboratory records curated from three published sources. To maximize data fidelity, the dataset underwent a standardized QA pipeline (unit harmonization, duplicate removal, consistency checks, and range screening against reported hydrate operating windows), with applicability-domain flags and outlier diagnostics retained for transparency. Model performance was assessed with grouped cross-validation by experiment to avoid leakage; the CNN–LSTM achieved the best accuracy (R 2 > 0.99; MSE ≈ 0.230 for methane recovery and 0.449 for CO 2 storage) and captured late-stage trends consistent with pore-scale transport limitations. A sensitivity analysis highlights the dominant influence of injection time and initial methane-hydrate saturation, in agreement with trends reported in the literature. No synthetic/augmented data were used; results therefore reflect measured behavior within the curated operating envelope. While the present dataset is modest, it is quality-controlled and literature-consistent; planned work will expand coverage (e.g., high-clay, high-salinity, multi-basin sediments) through new collaborations and prospective experiments to strengthen external validity and broaden the applicability domain. Overall, the AI proxy enables rapid, accurate screening of CH 4 /CO 2 replacement scenarios, reducing experimental burden and supporting data-driven decisions for carbon-neutral energy development and durable subsurface storage.