Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

An Integrated Framework Combining Computational Fluid Dynamics and Deep Learning for Hydrogen Flow Prediction

Karan, Michael LU (2025) MVKM01 20251
Department of Energy Sciences
Abstract
Hydrogen plays a pivotal role in the global energy transition, offering a clean and versatile energy
carrier for a sustainable future. However, ensuring its safe use in industrial facilities requires reli-
able prediction of dispersion from Venting Units (VUs). Traditional computational fluid dynamics
(CFD) simulations provide high-fidelity results but are too slow for fast safety assessment and
interactive engineering analysis. This thesis addresses these limitations by developing advanced
Deep Operator Network (DeepONet) models that deliver rapid, accurate predictions of hydro-
gen dispersion from VUs across a continuous parameter space. The proposed approach combines
multi-resolution Fourier feature encoding, residual network... (More)
Hydrogen plays a pivotal role in the global energy transition, offering a clean and versatile energy
carrier for a sustainable future. However, ensuring its safe use in industrial facilities requires reli-
able prediction of dispersion from Venting Units (VUs). Traditional computational fluid dynamics
(CFD) simulations provide high-fidelity results but are too slow for fast safety assessment and
interactive engineering analysis. This thesis addresses these limitations by developing advanced
Deep Operator Network (DeepONet) models that deliver rapid, accurate predictions of hydro-
gen dispersion from VUs across a continuous parameter space. The proposed approach combines
multi-resolution Fourier feature encoding, residual network connections, and novel asymmetric
gated fusion to overcome spectral bias and improve spatial accuracy. An automated pipeline trans-
forms high-fidelity CFD data from 168 parameter combinations into deployable AI models using
Latin Hypercube Sampling and systematic optimization. Systematic evaluation demonstrates that
the best DeepONet model achieves a velocity mean absolute error (MAE) of 0.167 m/s and a mass
fraction MAE of 0.000624, with correlation coefficients (R²) exceeding 0.999999. The method com-
presses data by a factor of 7,500 and reduces evaluation time from hours to milliseconds, enabling
instant 3D field visualization and interactive parameter exploration on standard hardware. This
work provides a practical, physically consistent tool for fast hydrogen safety assessment in indus-
trial environments. (Less)
Popular Abstract
Hydrogen is a promising clean energy source for a sustainable future, but safely handling it in industrial settings is a major challenge. Traditional simulations that predict how hydrogen spreads from venting units are accurate but too slow for real-time safety checks. This work introduces a fast, AI-based method that can predict hydrogen dispersion instantly and with high accuracy. By combining advanced neural network techniques and smart data processing, the model can visualize 3D hydrogen flow in milliseconds, making it possible for engineers to explore safety scenarios interactively. This approach provides a practical tool for ensuring hydrogen safety in industrial facilities while supporting the transition to cleaner energy.
Please use this url to cite or link to this publication:
author
Karan, Michael LU
supervisor
organization
course
MVKM01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Hydrogen Global energy transition Clean energy / sustainable energy Energy carrier Hydrogen dispersion Venting Units (VUs) Industrial safety Computational Fluid Dynamics (CFD) High-fidelity simulation Deep Operator Network (DeepONet) Rapid prediction / fast evaluation Multi-resolution Fourier feature encoding Residual network connections Asymmetric gated fusion Spectral bias Spatial accuracy Automated AI pipeline Latin Hypercube Sampling Systematic optimization Velocity MAE Mass fraction MAE Correlation coefficient (R²) Data compression Interactive 3D visualization Parameter exploration Physically consistent modeling Hydrogen safety assessment Industrial applications
report number
ISRN LUTMDN/TMPH-25/5660-SE
ISSN
0282-1990
language
English
id
9210648
date added to LUP
2025-08-28 09:29:07
date last changed
2025-08-28 09:29:07
@misc{9210648,
  abstract     = {{Hydrogen plays a pivotal role in the global energy transition, offering a clean and versatile energy
carrier for a sustainable future. However, ensuring its safe use in industrial facilities requires reli-
able prediction of dispersion from Venting Units (VUs). Traditional computational fluid dynamics
(CFD) simulations provide high-fidelity results but are too slow for fast safety assessment and
interactive engineering analysis. This thesis addresses these limitations by developing advanced
Deep Operator Network (DeepONet) models that deliver rapid, accurate predictions of hydro-
gen dispersion from VUs across a continuous parameter space. The proposed approach combines
multi-resolution Fourier feature encoding, residual network connections, and novel asymmetric
gated fusion to overcome spectral bias and improve spatial accuracy. An automated pipeline trans-
forms high-fidelity CFD data from 168 parameter combinations into deployable AI models using
Latin Hypercube Sampling and systematic optimization. Systematic evaluation demonstrates that
the best DeepONet model achieves a velocity mean absolute error (MAE) of 0.167 m/s and a mass
fraction MAE of 0.000624, with correlation coefficients (R²) exceeding 0.999999. The method com-
presses data by a factor of 7,500 and reduces evaluation time from hours to milliseconds, enabling
instant 3D field visualization and interactive parameter exploration on standard hardware. This
work provides a practical, physically consistent tool for fast hydrogen safety assessment in indus-
trial environments.}},
  author       = {{Karan, Michael}},
  issn         = {{0282-1990}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{An Integrated Framework Combining Computational Fluid Dynamics and Deep Learning for Hydrogen Flow Prediction}},
  year         = {{2025}},
}