Bayesian inference of light-gas dispersion from multi-fidelity data

Hydrogen is a promising alternative fuel due to its high energy density and carbon-free combustion. However, safe and widespread adoption requires effective hazard identification, assessment, and mitigation, including rapid hydrogen dispersion modeling for early risk mitigation. Data-driven modeling offers rapid predictive capabilities; however, such approaches are costly due to the need for vast amounts of high-quality or high-fidelity (HF) data. To this end, multi-fidelity (MF) modeling can fuse scarce HF data with abundant low-fidelity (LF) data for rapid HF predictions. In this work, we used helium as a surrogate light gas of hydrogen and developed a Bayesian regression-based MF model to predict the gas concentration in a semi-confined dispersion chamber at multiple locations for various operating conditions. The MF model predicts distributions of potential concentration values, allowing for uncertainty quantification and a statistical assessment of leakage risk. Both experimental and numerical approaches are used to generate datasets used in the current study. The experiment is designated as the HF model, serving as the ground truth with fewer data points. The numerical approach employing computational fluid dynamics (CFD) is designated as the LF model, with more data sets generated at a relatively lower cost. The MF model was trained on all LF data and a subset of the HF data. Good predictive performance is observed on the testing HF data, indicated by an R² value of 0.9, particularly for physical locations away from the leakage point. This study provides new pathways to rapidly assess hydrogen safety and mitigate early-stage risk with minimal HF data.