The problem
Emissions limits on NOx and CO, and the push toward alternative and sustainable fuels, make emissions prediction a first-order constraint in gas-turbine design. CFD with tabulated chemistry is the most trusted predictive tool, but a single RANS simulation of a combustor costs hours to days on hundreds of cores, far too slow to sweep a design space.
My work attacks this from both sides: physics-based reduced-order models that keep detailed chemistry at a fraction of the cost, and machine learning that builds, corrects, and enriches those models using scarce high-fidelity data.
1. High-fidelity combustor CFD
I run both RANS and large-eddy simulations (LES) of swirl-stabilized, liquid-fueled gas-turbine combustors, capturing the full chain from fuel injection through spray breakup, evaporation, mixing, and turbulent combustion. The RANS campaign of a Jet-A can combustor in CONVERGE drives the reactor-network work; the LES provide high-resolution reference fields, like the one in the video above.
- Turbulence and combustion: RANS with standard k-ε closure; flamelet-generated manifold (FGM) chemistry with a presumed β-PDF for turbulence–chemistry interaction, and a separate NO progress variable for slow NOx chemistry.
- Sprays: Lagrangian discrete-droplet tracking, Taylor Analogy Breakup (TAB) for primary and secondary breakup, Frossling evaporation, and Rosin-Rammler size distribution.
- Numerics: adaptive mesh refinement from 3.5 mm down to 0.44 mm in high-gradient regions, about 3.2 million cells, roughly 10 hours on 256 cores.
2. Liquid-Fueled Reactor Network (LFRN)
First author · Combustion and Flame, 2026
Reactor networks replace millions of CFD cells with a small graph of idealized chemical reactors. Existing methods assume gaseous fuel. I developed what is, to our knowledge, the first truly coupled reactor-network framework for liquid-fueled combustors.
- Data-driven zoning: k-means clustering of CFD fields defines reactors that map onto physical regimes: spray breakup, mixing, two flame zones, two recirculation zones, and post-flame.
- New spray reactors: an evaporator/breakup reactor and a mixer reactor that model droplet breakup, heating, and evaporation directly inside the network, alongside conventional perfectly stirred and plug-flow reactors.
- Detailed chemistry: Cantera with the HyChem Jet-A mechanism (119 species, 841 reactions), resolving thermal, prompt, and N2O-intermediate NOx pathways.
- Validation: against CFD across inlet air temperatures of 300–500 K and multiple fuel flow rates. The LFRN substantially reduces NOx error compared with a gaseous reactor network while keeping outlet temperature accurate, and converges toward CFD as clusters are added, with diminishing returns beyond 9.
- Speed: O(10) s on a single CPU core. The 7-cluster network is about 2,300× faster than the 256-core CFD; the 11-cluster network finishes in 35.8 s.
3. Reinforcement learning for automated model construction
First author · Proceedings of the Combustion Institute, 2026
Building a good reactor network has traditionally required expert judgment. I framed network construction as a sequential decision problem and trained reinforcement-learning agents to optimize how CFD fields are clustered into reactors. The RL-built networks predict lean blowout better than manually constructed ones.
4. Multi-fidelity Gaussian-process surrogates
First author · In preparation
Reactor networks are fast but biased; CFD is accurate but scarce. I asked: given a fixed CFD budget, does fusing abundant reactor-network data with scarce CFD through a multi-fidelity Gaussian process give a better emissions surrogate, and for which quantities?
- Setup: two inputs (fuel mass flow, air inlet temperature) and seven outputs (outlet temperature and NO, NO2, OH, CO, CO2, H2O), using 35 CFD points and 400 LFRN points. A high-fidelity GP is compared with a nonlinear autoregressive multi-fidelity GP (NARGP), averaged over 30 random splits at each of nine training fractions.
- Data efficiency: with only 7 CFD points, the multi-fidelity GP matches the accuracy a standard GP needs 14 for, lifting mean R² from 0.75 to 0.92 (outlet temperature 0.87 → 0.98, NO2 0.36 → 0.75).
- Negative transfer: the benefit is species-dependent. Because the LFRN's CO prediction is biased, fusing it degrades the CO surrogate. The paper gives practical guidance on when cheap low-fidelity data is worth fusing.
5. Diffusion models for flow-field reconstruction
Ongoing PhD project · In progress
Reactor networks give zone-averaged answers. One of the final projects of my PhD uses a diffusion model to reconstruct high-resolution LES flow and species fields from reactor-network data, recovering spatially resolved detail from a model that runs at a small fraction of LES cost.
6. HPC pipelines, neural networks & chemistry tabulation
- Parallel, automated Python pipelines on LSU's SuperMike-3 cluster (SLURM, MPI) linking simulation, data extraction, model construction, and ML training.
- Large parametric NOx and lean-blowout campaigns.
- Neural networks that predict NOx from reduced-order model outputs.
- Flamelet-generated manifold pipelines with OpenSMOKE++ that compress detailed chemistry into low-dimensional lookup tables.
Tools
Service & teaching
- Peer reviewer for Combustion and Flame and the Journal of Energy Resources Technology.
- Teaching assistant for Thermodynamics, Fluid Mechanics, and Heat Transfer, mentoring undergraduates on research and senior design.