BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.4.0.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:223@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20260831T000000 DTEND;TZID=Asia/Kolkata:20260831T000000 DTSTAMP:20260824T141737Z URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-defense-cds-data- driven-and-low-precision-methods-for-improving-and-accelerating-computatio nal-fluid-dynamics-solvers/ SUMMARY:M.Tech Research Thesis Defense: CDS: "Data-driven and low-precision methods for improving and accelerating computational fluid dynamics solve rs" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research Thesis Defense\n\n\n\nSpeaker : Mr. Surya Datta Sudhakar\nS.R. Number : 06 -18-01-10-22-24-1-24309\nTitle: "Data-driven and low-precision methods for improving and accelerating computational fluid dynamics solvers."\nThesis examiner : Dr. Rishita Das\, Aerospace Engineering\, IISc\nResearch Super visor: Dr. Konduri Aditya\nDate &\; Time : August 31\, 2026\, 10:00 AM\ nVenue : # 202 CDS Classroom\n\n\n\nABSTRACT\nAdvancing simulation of turb ulent and reacting flows poses two distinct challenges\, the accurate repr esentation of unresolved physical processes and the efficient utilization of the modern computing hardware. This thesis addresses both directions th rough data-driven subgrid-scale modeling and low-precision solver developm ent.\n\nThe first part of this thesis investigates the transferability of data-driven subgrid-scale closures for passive scalar transport in non-equ ilibrium turbulence\, motivated by the central importance of passive scala r mixing in a wide range of engineering and environmental flows\, includin g 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 an d impractical. This study considers decaying two-dimensional turbulence wi th passive scalar transport as a stringent testbed due to its non-stationa ry dynamics\, evolving spectra\, intermittency\, and strong nonlocal inter actions across scales.\n\nNeural-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 exami ned through selective fine-tuning of pretrained models and compared agains t conventional closure models\, demonstrating the superior adaptability of learned closures. The results reveal a strong directional asymmetry\, wit h models trained at higher Schmidt numbers transferring more effectively t o lower Schmidt regimes than the reverse. Layer-wise analysis shows that p redictive improvements arise primarily from adapting shallow network layer s\, while modifications to deeper layers often reduce performance. Loss la ndscape analysis further provides insight into the optimization behavior u nderlying successful transfer. Spectral analysis shows that the fine-scale behavior of the learned scalar closures exhibits greater universality acr oss Schmidt numbers\, whereas large-scale closure behavior remains more re gime-dependent. These findings establish a computationally efficient and p hysically interpretable framework for the a priori development and assess ment of transferable machine-learned scalar closures in large-eddy simulat ion.\n\nThe second part of this thesis investigates a low-precision framew ork for chemically reacting flow simulations through multiple implementati ons developed for complementary objectives. A Python implementation is emp loyed to assess the numerical feasibility of FP16 computations using softw are-based emulation together with a dynamic scaling strategy. Building upo n these accuracy studies\, performance-oriented implementations are develo ped 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 nati ve FP16 arithmetic units\, while the GPU implementation uses FP16 storage and data communication together with predominantly FP32 arithmetic and sel ective higher-precision treatment of numerically sensitive operations. Thi s design is motivated by the extreme dynamic range of species concentratio ns\, reaction rates\, and thermodynamic variables in chemically reacting f lows\, for which operations such as subtraction\, division\, and accumulat ion are particularly susceptible to precision loss in native FP16 arithmet ic. The framework incorporates adaptive scaling and exponent bookkeeping t o preserve numerical robustness while reducing memory footprint and commun ication cost.\n\nOverall\, this thesis investigates two complementary stra tegies for improving computational fluid dynamics solvers by developing tr ansferable machine-learned closures for improved large-eddy simulation mod eling and the development of a reduced-precision solver framework that low ers memory and data transfer costs while maintaining numerical accuracy.\n \n\n\nALL ARE WELCOME CATEGORIES:Events,Thesis Defense END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20250831T000000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR