Machine learning assisted Direct Numerical Simulation of non-premixed Ammonia Combustion in inert porous material

  • Workgroup:Simulation of Reacting Thermo-Fluid Systems
  • Type:Ba/Ma
  • Date:immediately
  • Supervisor:

    M.Sc. Rishabh Puri

  • Background knowlegde:

    Students of chemical engineering/process engineering (or similar) interested in numerical work. Knowledge of numerical flow simulation, fluid mechanics , machine learning models and programming skills (e.g. C/C++, Python, Matlab or similar) may help you get started but is not required.

  • Location: CS
    Ammonia offers a promising possibility as a carbon-free energy carrier. Although it can be easily transported using the existing infrastructure and stored for future energy needs, combustion of ammonia poses considerable challenges. The three major challenges are its low burning velocity compared to hydrocarbons resulting in poor flame stability, extremely high levels of nitrogen oxide formation and high toxicity even at trace levels. To tackle these problems, a model burner is proposed using the heat transfer properties of inert porous material.


    Project description:
    Various test cases are defined using two-dimensional burner configurations (see Fig. 1) for non-premixed combustion of ammonia and the same are to be simulated by performing reactive Direct Numerical Simulations (DNS). Simulations are initialised with different inlet temperatures, stoichiometric conditions, fuel blends and flow velocities. Additionally, the boundary conditions on the burner are varied. Given the unsteady nature of the simulations, it takes some time before the flame is stabilised. There may be some burner configurations, for which a stable flame is not possible. Similarly, the concentration of the nitrogen oxides and residual ammonia in exhaust (ammonia slip) will vary with simulation setup. The aim of this work is to study the flame properties and exhaust products for different burner configurations. Simulating a large number of burner configurations using DNS is time consuming and costly. To reduce the number of useful configurations, machine learning based models such as Graph Neural Networks (GNNs), enforced with constraints from combustion chemistry, need to be implemented. As an application example of the GNN approach, Fig. 2 shows a computational grid for flow around a cylinder, that is converted into a graph by extracting the graph attributes from the grid.


    Figure 1: Two dimensional burner with structures resembling porous material:


    Figure 2: Extraction of attributes from a computational grid to generate graphs for training GNNs:



    • Conduct a literature review to gain knowledge of reacting flows and the chemistry of the ammonia combustion.
    • Familiarise with the basics of numerical methods, computational fluid dynamics and OpenFOAM
    • Implement Physics-Informed and Chemistry-Assisted Graph Neural Networks (PICA-GNN)
    • Validate machine learning models using DNS results
    • Define suitable test cases to be simulated using DNS on OpenFOAM
    • Document the major findings in a BSc/MSc thesis und present the final results


    Learning Outcomes:

    • Apply numerical and machine learning methods to engineering problems
    • Conduct simulations of reacting flows with open-source tools
    • Develop tools for scientific programming



    Prof. Dr. Oliver T. Stein