Active Learning for Transport Property Prediction in CO ₂ –Hydrocarbon Systems: A Multifidelity Approach Integrating Molecular Dynamics and Experiments

Experimental studies of petroleum-related systems often require extreme temperature and pressure conditions, increasing costs, complexity, and safety risks of laboratory work. More targeted and resource-efficient approaches are therefore essential. Active learning combined with machine learning offers a powerful means to strategically select informative experiments and reduce data demands for predictive modeling. Here, we introduce a multifidelity active learning framework based on Gaussian process regressions to predict viscosity and mutual diffusivity in individual mixtures containing CO2, hydrocarbons, and polar solvents. It combines molecular dynamics simulations as a low-fidelity, cost-effective source of physically meaningful data with high-fidelity experiments compiled from the literature. Accurate models were obtained for each mixture using only 5–9 experiments and 7–13 simulation data points, compared to 22–75 experiments in conventional workflows. Prediction errors remained below 6% across broad thermodynamic conditions (1–700 bar, 278–423 K) and compositions. These results demonstrate that multifidelity active learning supported by molecular dynamics enables reliable property prediction with minimal experimental input, paving the way to streamline workflows in practical petroleum and low-carbon technologies and to mitigate risks in extreme-condition operations.