Sushrut Kumar

• Developed numerical methods and algorithms in a large‑scale lab’s codebase for turbulent flow simulations with moving dirac delta discontinuity. • Independently ported entire codebase to multi-GPU platform. Achieved over 90% parallel efficiency using CUDA graphs, CUDA aware P2P communication and collectives to solve massive systems(≈0.5 billion simulation points). • Packaged simplified codebase to ImmerseFlow++ using CUDA and C++. • My PhD work resulted in a software that is 1000X (10X numerical, 2X algorithmic improvement and 50X GPU acceleration) faster than earlier version and enables our ability to solve and understand problems like biolocomotion, heart valves etc. • Developed deep learning architectures by fusing existing numerical methods and graph neural networks for fast PDE solutions. • Developed a distributed data processing pipeline for physics and data‑driven analysis of terabyte‑scale simulation data (≈30TB) and contributed to 3 journal papers.

Teaching Assistant for 1. Applied Computational Methods for Aerodynamics (Graduate Level) 2. Numerical Methods by Prof. Rajat Mittal, Fall 2022 (Graduate Level) 3. Heat Transfer by Prof. Charles Meneveau, Spring 2022 (Undergraduate Level) Responsibilities include preparing and correcting assignments, holding office hours, and conducting weekly problem-solving sessions

Developing geometric deep learning architectures and inference pipeline for spatio-temporal predictions of PDE solutions within digital twin environment

• Worked on entire product lifecycle from conceptualization to development and deployment. • Developed deep learning architectures using physics-informed machine models to create physics-aware digital twins providing 5000X in performing calculations than SOTA. • Independently developed end-to-end pipeline from data generation, model development, multi-GPU training, and model deployment on cloud infrastructure. • Designed systems to integrate live sensor data feed for real-time inference within Nvidia Omniverse. • Contributed to 5 patents

Developed and validated a CNN autoencoder model using TensorFlow for fast computation of Navier-Stokes and energy PDE solutions in channel flows with thermally complex sinusoidal boundaries.

• Conducted quantitative analysis on fluid mechanics data using regression and neural network models, achieving a 340% increase in test scores. • Developed an aerodynamic shape optimization framework with genetic algorithms and ML techniques, resulting in a 71% optimization rate. • Investigated turbulence attenuation through large eddy simulations and statistical tools.

• Developed a Python code to model active matter locomotion in viscous fluids using stochastic rotation dynamics, optimizing code from O(N^D K) to O(K) along with LLVM compilation and achieving 40X speedup. • Validated models against Hagen–Poiseuille, Couette flow, and flow with bacterial microsquirmers.

Focus - Computational Fluid Dynamics, Machine Learning & High Performance Computing

Focus - Application of Soft Computing in Fluid Dynamics