Logo image
Mesh-aware multi-fidelity surrogate modelling for data-efficient hydrogen dispersion field prediction
Doctoral Thesis   Open access

Mesh-aware multi-fidelity surrogate modelling for data-efficient hydrogen dispersion field prediction

Xiaoyang Luan
University of Surrey
Doctor of Philosophy (PhD), University of Surrey
30/06/2026
DOI:
https://doi.org/10.15126/thesis.902113

Abstract

Hydrogen Dispersion Process Safety Chemical Engineering Computational Fluid Dynamics Machine Learning
Hydrogen refuelling stations (HRS) sit at the interface between high-pressure hydrogen infrastructure and public environments, where accidental releases can form flammable clouds. Predictive modelling is hindered by the cost and mesh dependence of computational fluid dynamics (CFD) simulations, which become computationally intensive at high-fidelity (HF) resolution. This thesis develops a data-efficient, uncertainty-aware workflow for hydrogen dispersion prediction that treats mesh resolution as the primary fidelity axis.</p><p>First, credibility of Reynolds-averaged Navier–Stokes (RANS) simulation for under-expanded hydrogen jets is strengthened by calibrating key closure coefficients in the standard <span style="color: black;"> </span> model. Gaussian process (GP) surrogate modelling enables global (Sobol) sensitivity analysis to rank coefficients and supports targeted calibration against experimental datasets, producing a defensible CFD configuration for subsequent data generation.</p><p>Second, to reduce reliance on expensive fine-mesh runs, a mesh-aware multi-fidelity Gaussian process (MAMF-GP) is proposed. Using the mesh ratio as a continuous input, its covariance embeds mesh-convergence structure by separating a mesh-independent physical trend from primary and higher-order discretisation-error channels whose influence decays with refinement. Convergence rates are learned via marginal-likelihood training, improving accuracy and delivering coherent uncertainty quantification under limited HF budgets.</p><p>Third, the mesh-aware strategy is extended to full-field prediction through reduced-order modelling. A multi-level residual proper orthogonal decomposition (MLR-POD), a proper orthogonal decomposition (POD) variant, constructs an HF-consistent reduced basis from nested meshes by combining coarse modes with successive refinement residual modes. MAMF-GP then performs mesh-aware latent-space regression of reduced coefficients, enabling accurate reconstruction of HF concentration fields and prediction of the spatial extent of lower flammability limit region. </p><p>Overall, the thesis couples CFD calibration with mesh-aware multi-fidelity surrogate and reduced-order modelling to enable faster and more reliable dispersion prediction for HRS safety assessment and related mesh-refined simulations.
pdf
Thesis_final version_for_submission_Xiaoyang Luan7.38 MBDownloadView
Open Access CC BY-NC-SA V4.0

Metrics

1 Record Views

Details

Logo image

Usage Policy