DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis Defense
Speaker : Mr. Surya Datta Sudhakar
S.R. Number : 06-18-01-10-22-24-1-24309
Title: “Data-driven and low-precision methods for improving and accelerating computational fluid dynamics solvers.”
Thesis examiner : Dr. Rishita Das, Aerospace Engineering, IISc
Research Supervisor: Dr. Konduri Aditya
Date & Time : August 31, 2026, 10:00 AM
Venue : # 202 CDS Classroom
ABSTRACT
Advancing simulation of turbulent and reacting flows poses two distinct challenges, the accurate representation of unresolved physical processes and the efficient utilization of the modern computing hardware. This thesis addresses both directions through data-driven subgrid-scale modeling and low-precision solver development.
The first part of this thesis investigates the transferability of data-driven subgrid-scale closures for passive scalar transport in non-equilibrium turbulence, motivated by the central importance of passive scalar mixing in a wide range of engineering and environmental flows, including combustion, atmospheric transport, pollutant dispersion, and heat and mass transfer. Accurate modeling of unresolved scalar transport remains a major challenge in large-eddy simulation, while training separate neural-network closures for every flow condition is computationally expensive and impractical. This study considers decaying two-dimensional turbulence with passive scalar transport as a stringent testbed due to its non-stationary dynamics, evolving spectra, intermittency, and strong nonlocal interactions across scales.
Neural-network-based closure models are trained to predict exact subgrid-scale closure targets from filtered high-fidelity simulation data and are evaluated through a priori analysis against the corresponding filtered DNS data, with the Schmidt number governing scalar mixing behavior. Transfer learning across Schmidt number regimes is examined through selective fine-tuning of pretrained models and compared against conventional closure models, demonstrating the superior adaptability of learned closures. The results reveal a strong directional asymmetry, with models trained at higher Schmidt numbers transferring more effectively to lower Schmidt regimes than the reverse. Layer-wise analysis shows that predictive improvements arise primarily from adapting shallow network layers, while modifications to deeper layers often reduce performance. Loss landscape analysis further provides insight into the optimization behavior underlying successful transfer. Spectral analysis shows that the fine-scale behavior of the learned scalar closures exhibits greater universality across Schmidt numbers, whereas large-scale closure behavior remains more regime-dependent. These findings establish a computationally efficient and physically interpretable framework for the a priori development and assessment of transferable machine-learned scalar closures in large-eddy simulation.
The second part of this thesis investigates a low-precision framework for chemically reacting flow simulations through multiple implementations developed for complementary objectives. A Python implementation is employed to assess the numerical feasibility of FP16 computations using software-based emulation together with a dynamic scaling strategy. Building upon these accuracy studies, performance-oriented implementations are developed in C++ for CPUs and C++/CUDA for GPUs. The CPU implementation employs FP16 storage with arithmetic promoted to FP32 owing to the absence of native FP16 arithmetic units, while the GPU implementation uses FP16 storage and data communication together with predominantly FP32 arithmetic and selective higher-precision treatment of numerically sensitive operations. This design is motivated by the extreme dynamic range of species concentrations, reaction rates, and thermodynamic variables in chemically reacting flows, for which operations such as subtraction, division, and accumulation are particularly susceptible to precision loss in native FP16 arithmetic. The framework incorporates adaptive scaling and exponent bookkeeping to preserve numerical robustness while reducing memory footprint and communication cost.
Overall, this thesis investigates two complementary strategies for improving computational fluid dynamics solvers by developing transferable machine-learned closures for improved large-eddy simulation modeling and the development of a reduced-precision solver framework that lowers memory and data transfer costs while maintaining numerical accuracy.
ALL ARE WELCOME
