10–12 Nov 2026
Arcotel Wimberger
Europe/Vienna timezone

Data Driven Machine Learning Approach for 3D Flow Prediction in Hydrogen Burners

12 Nov 2026, 11:00
20m
Room 1

Room 1

Oral Presentation Digitalisation, Artificial Intelligence & Simulation Digitalisation & Artificial Intelligence

Speaker

Luca Juris (Chair for Digital Additive Production DAP - RWTH Aachen)

Description

Hydrogen burner design requires accurate prediction of three-dimensional flow and mixing fields, since local hydrogen distribution strongly influences flame stability, flashback risk, and emission behavior. Additive manufacturing enables the realization of complex burner concepts with tailored injection layouts, small-scale fuel inlets, and compact mixing geometries that would be difficult or impossible to manufacture using conventional processes. At the same time, this increased design freedom requires efficient simulation-based methods to evaluate large design spaces while respecting manufacturing constraints. In this work, the burner geometry is therefore represented by a parametrized, manufacturing-aware model in which relevant additive manufacturing restrictions are embedded directly in the admissible design space.

The proposed approach uses a MeshGraphNet architecture trained on CFD-generated flow fields from OpenFOAM simulations. Unstructured CFD meshes are transformed into graph representations, where nodes contain local flow quantities, spatial information, and boundary encodings, while edges describe geometric relations between neighboring mesh points. The model predicts incremental updates of velocity, pressure, and species mass fractions and is applied autoregressively for multi-step rollout prediction. To improve physical plausibility, the data-driven training objective is extended with a lightweight species-closure constraint that penalizes violations of species consistency.

Results (Fig. 1) on unseen burner geometries in the parameter space show stable rollout behavior and good agreement with CFD reference data for velocity and pressure. The relative error increases moderately over the rollout horizon. The species constraint improves the plausibility and spatial symmetry of the hydrogen field without requiring additional CFD data or expensive residual evaluations. However, hydrogen mass fraction remains the most challenging prediction target, especially in low-concentration regions where small local deviations are not fully captured by global error metrics. These findings indicate that species closure alone is not sufficient to achieve CFD-level fidelity for hydrogen mixing. Ongoing work is therefore focused on integrating stronger PDE-based loss terms derived from the governing transport equations to improve accuracy in critical low-concentration zones and enhance the physical consistency of long-term rollouts.
Fig 1: Qualitative comparison on a representative unseen test geometry. Mid plane slices show ground truth CFD fields (bottom), and MGN predictions (top), for velocity magnitude (left), pressure (middle), and hydrogen mass fraction (right). The model reproduces the main flow structures and mixing layer development, while errors remain localized near regions of high gradients or fine scale values.

Overall, the study demonstrates that constrained MeshGraphNets are a promising surrogate modeling approach for accelerating hydrogen burner design. Even before reaching standalone CFD accuracy, the model can provide informed initial conditions for CFD simulations, and serve as a foundation for future physics-enhanced optimization workflows.

Speaker Country Germany
Would you like to publish your paper in the special issue of BHM "Berg- und Hüttenmännische Monatshefte" Yes

Authors

Luca Juris (Chair for Digital Additive Production DAP - RWTH Aachen) Mr Kaan Atak Johannes Henrich Schleifenbaum (Chair for Digital Additive Production DAP - RWTH Aachen)

Presentation materials

There are no materials yet.