Modeling CO2 leak and dispersion from injection wells and surface facilities using machine learning

Abstract

This study presents a novel machine learning–based approach for real-time predictions of CO₂ leak and dispersion from injection wells and surface facilities. A comprehensive dataset was generated using PHAST simulations under worst-case scenarios, specifically targeting two critical CO₂ concentration thresholds, 30,000 ppm (ACGIH STEL) and 40,000 ppm (IDLH). Key operational parameters, including operating temperature (0–100 °C), operating pressure (1–250 bar), and leak diameter (1–12 inches), were systematically varied while maintaining other parameters fixed at conservative values, and considering two atmospheric stability classes (F and D) with corresponding surface wind speeds (1.5 m s⁻¹ and 5 m s⁻¹) respectively. Multiple machine learning models were trained on this high-fidelity dataset, and the model performances were evaluated. The Artificial Neural Network (ANN) model emerged as the top performer, achieving high predictive accuracy (R² ≈ 0.9996 on test data) with minimal error. Extensive diagnostic analyses, including residual, cumulative error, and leverage evaluations, confirmed the model's robustness and generalizability. The final predictive model was integrated into an interactive application, enabling rapid hazard assessment, and offering a significant reduction in computational cost compared to traditional Computational Fluid Dynamics (CFD) methods. This work provides a scalable framework for enhancing emergency response and process safety in CO₂-intensive industrial environments.