DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
Ph.D. Thesis Colloquium
Speaker: Mr. Priyabrat Dash
S.R. Number: 06-18-01-10-12-20-2-19095
Title: Toward machine learning-accelerated simulations of cavity-stabilized flames
Research Supervisor: Dr. Konduri Aditya
Date & Time : April 20, 2026 (Monday), 11:00 AM
Venue : #102, CDS Seminar Hall
ABSTRACT
Direct numerical simulations (DNS) provide a rigorous framework for resolving the multiscale interactions governing turbulent reacting flows in practical combustors. By capturing all dynamically relevant spatio-temporal scales, DNS complements experimental approaches by providing access to flow and thermochemical quantities that are difficult to measure at the required spatio-temporal resolution. Cavity flameholders, essential to gas turbines and ram/scramjet engines, represent one such configuration where this capability is especially valuable. These devices generate low-speed recirculation zones that continuously entrain heat and radicals, facilitating flame stabilization. Despite its unparalleled detail, the prohibitive computational cost of DNS limits its adoption for complex configurations. This thesis leverages the data-rich nature of DNS to develop machine learning frameworks that can accelerate simulation workflows, augment experimental diagnostics, and yield deeper physical insight into the dynamics of cavity-stabilized combustion.
To address the computational cost of DNS, this work first develops a physics-guided self-supervised super-resolution (SR) methodology to reconstruct fine-scale structures in isotropic turbulence from coarse-grained flow fields. By treating SR as a state estimation problem, this approach does not require paired high-resolution training data, making it well-suited for flow configurations where fully resolved simulations are prohibitive. Separately, a graph neural network (GNN) based SR framework is developed for turbulent reacting flows on complex meshes, including structured non-uniform and unstructured configurations. Leveraging message passing layers, this approach enables accurate reconstruction of unresolved flow features across geometrically complex domains, extending the applicability of data-driven SR to practical combustion configurations.
The mesh-native nature of GNNs established in the SR framework naturally extends to other modeling challenges in reacting flow simulations. A GNN-based closure has been developed to predict filtered species production rates from filtered thermochemical scalars on non-uniform meshes, with a priori evaluations demonstrating strong generalization across varying hydrogen blend compositions and filter widths. To further reduce/alleviate the complexity of detailed chemical kinetics, two GNN-based mechanism reduction formulations are introduced, GNN-SM and GNN-AE, leveraging surrogate-assisted and autoencoder-based strategies respectively to identify kinetically dominant pathways, achieving up to 95% reduction in species and reactions while maintaining predictive accuracy.
The mesh-agnostic nature of GNNs established in the SR framework naturally extends to other modeling challenges in reacting flow simulations. A GNN-based closure has been developed to predict filtered species production rates from filtered thermochemical scalars on non-uniform meshes. A priori evaluation has demonstrated strong generalization across varying fuel blend compositions and filter widths. To reduce the complexity of detailed chemical kinetics, two GNN-based mechanism reduction formulations are introduced, GNN-SM and GNN-AE, leveraging surrogate-assisted and autoencoder-based strategies respectively to identify kinetically dominant pathways. Up to 95% reduction in species and reactions has been observed, while maintaining predictive accuracy.
DNS data is further used to augment experimental diagnostics in trapped vortex combustors. Vision transformer (ViT) models are trained on simulation data with synthetic noise to infer velocity fields from planar laser-induced fluorescence (PLIF) measurements, avoiding the intrusiveness of particle image velocimetry. These models demonstrate accurate reconstruction of large-scale vortical structures, enabling data fusion and digital twin development where intrusive diagnostics are impractical.
Finally, DNS is employed to advance our physical understanding of cavity-stabilized combustion. Massively parallel compressible flow simulations of lean premixed ethylene-air flames in a backward-facing step reveal how the primary recirculation zones transport radicals and alter dominant chemical pathways to support flame stabilization. These investigations are extended to the HIFiRE 2 supersonic combustor, where dual-mode and scram-mode operations are analyzed to characterize flame stabilization regimes and the evolution of chemical pathways in non-premixed supersonic flames. Collectively, this work integrates high-fidelity DNS with geometric-aware machine learning frameworks, establishing a foundation for accelerated simulations, improved diagnostics, and reduced-order modeling in cavity-stabilized reacting flows.
ALL ARE WELCOME
