Department of Computational and Data Sciences
Department Seminar
Speaker : Dr. Shubhangi Bansude, Assistant Professor, Mechanical Engineering Department, IIT Gandhinagar
Title : Physics-Guided Deep Learning for Large-Eddy Simulation of Turbulent Combustion.
Date & Time : June 29th, 2026 (Monday), 11:30 AM
Venue : # 102, CDS Seminar Hall
ABSTRACT:
The transition toward carbon-free propulsion and power systems demands advancedcomputational models capable of accurately predicting and optimizing both existing and emerging combustion technologies. Large-eddy simulation (LES) offers a promising approach, resolving large-scale spatiotemporal features and capturing transient phenomena such as flame dynamics and instabilities. However, LES still faces major challenges: modeling the effects of unresolved subgrid scale (SGS) turbulence, managing the high computational cost of finite-rate chemical kinetics, and accurately representing turbulence-chemistry interactions.
This talk presents a physics guided, data-driven approach that leverages deep learning to address these challenges. It will first discuss a neural ordinary differential equation (NODE) framework developed to accelerate the integration of stiff chemical kinetics within the transported filtered density function (FDF) method, where physical constraints are embedded as a soft penalty to improve adherence to physical laws and generalizability. Results will be presented for this framework across multiple fuels – hydrogen, methane, and ammonia – in a pairwise mixing stirred reactor (PMSR) spanning a range of mixing timescales, demonstrating that the NODE substantially reduces numerical stiffness and achieves an order-of-magnitude speedup over direct integration of detailed kinetics. The talk will then turn to a complementary, lower-fidelity alternative: a moments-based method that condenses the chemistry description to a small set of characteristic variables and incorporates SGS statistics through a presumed FDF. Within this framework, a deep neural network (DNN) trained to predict the FDF of the mixture fraction will be presented, along with results showing how this DNN-FDF model, evaluated in LES of variable-density 3D mixing layers, improves accuracy over the conventional beta-FDF model in reproducing direct numerical simulation (DNS) results. The talk will conclude with a discussion of emerging opportunities and open challenges in embedding physics-informed machine learning within multi-scale simulations of turbulent reacting flows.
BIOGRAPHY:
Dr. Shubhangi Bansude is an Assistant Professor in the Mechanical Engineering Department at IIT Gandhinagar, where she is PI of the Computational Combustion and Machine Learning (CCML) Lab. She is a researcher in turbulent combustion modeling with specialization in scientific machine learning and combustion chemical kinetics. In particular, her work involves developing physics-guided, data-driven approaches for turbulent combustion modeling and applying them to the simulation of full-scale energy systems, with an emphasis on large-eddy simulation (LES). Previously, she was a Postdoctoral Researcher at Argonne National Laboratory. She completed her PhD at the University of Connecticut (UConn) in 2023, during which she served as a Graduate Teaching Fellow and was awarded the prestigious General Electric Graduate Fellowship for Innovation. Prior to her doctoral studies, she worked as an Application Engineer at Siemens PLM Software contributing to advanced computational fluid dynamics (CFD) solutions for the automotive and OEM industries. She obtained her BTech in Mechanical Engineering from IIT Gandhinagar in 2014.
Host Faculty: Dr. Konduri Aditya
ALL ARE WELCOME
