Research Projects

Hydrogen (H₂) and ammonia (NH₃), when produced from renewable or low-carbon sources associated with carbon capture and storage, are regarded as promising clean fuels for mitigating climate change (Gopalakrishnan et al. 2024; Mohammed et al. 2024). However, hydrogen poses challenges such as flashback, thermo-diffusive instability, and difficulties in storage and transport. Ammonia, though easier to handle, is less reactive, hard to ignite, and leads to high NOx emissions (Coulon et al. 2023; Yang et al. 2022). Blending NH₃ with H₂ enhances ignition but still results in significant NOx (Bioche et al. 2021; Tian et al. 2023). To control NOx and enable fuel flexible operation with increased efficiency, sequential combustors are used in gas turbines (Ansaldo Energia GT36) (Aditya et al. 2019; Gopalakrishnan et al. 2024; Gruber et al. 2021). Such systems have two combustion stages axially separated by a zone for injecting dilution air and additional fuel. The first stage is flame-propagation stabilized, while the second relies on auto-ignition in vitiated mixtures (Aditya et al. 2019; Heggset et al. 2024) . Understanding auto-ignition of NH₃-H₂ and optimizing mixture fractions for stable, low NOx operation is the key focus. While premixed flames are well studied (Bioche et al. 2021; Tian et al. 2023), auto-ignition remains challenging due to sensitivities to variable fuel reactivity, turbulence-chemistry interaction, reliable reaction mechanisms and NOx pathways (Chi, Han, and Thévenin 2023; Yang et al. 2022). Direct Numerical Simulation (DNS) resolves all scales but is costly. Whereas, Large-Eddy Simulations (LES) are less computationally intense, resolves large scale eddies while using sub-grid scale (SGS) models for small scale ones. Major differences in ignition delay, heat release rates, and NOx formation are expected when the simulations are not fully resolved (Zhou et al. 2022). The reliability of LES results depends on SGS models. Therefore, the primary objective of this work is to investigate auto-ignition flames of NH₃-H₂ blends in sequential combustor using LES with machine learning (ML) for improved model fidelity. SGS models provide closures for turbulence stress, scalar mixing, and filtered reaction rates for the unresolved scales (Zhou et al. 2022). These are typically calibrated using experimental or DNS data and are well established for CH₄ (Colin et al. 2000; Charlette et al. 2002) and H₂ (Domingo et al. 2008), but not for NH₃-H₂ blends. Therefore, this work aims to learn the SGS closures for auto-ignition of NH₃-H₂ fuel blends in sequential combustors by training neural networks (NN) on DNS data. The resulting ML-based SGS models will be incorporated into LES solver to enhance the accuracy of LES simulations. This LES-ML framework can further be used to study modes of combustion of NH₃-H₂ in sequential combustor, flashback and blow-out, thermo-acoustic instability, fuel split and air split ratio optimization to minimize NOx.