As a Simulation Intelligence Scientist, I

• Research, Design and implement production-ready graph and point-cloud surrogate neural operator models in JAX and PyTorch, along with robust data pipelines and preprocessing workflows to support reliable training, evaluation, and deployment.
• Investigate effective geometry representation in point clouds and global-interaction mechanisms in graph neural networks, including approaches to incorporate physical symmetries for steady-state surrogate modeling and improved generalization across diverse CAD geometries.
• Develop subsampling and subgraph/partitioning strategies to scale learning and inference to large CAE datasets, reducing memory bottlenecks and enabling efficient processing of high-resolution simulations.
• Collaborate with cross-functional engineering teams and external research partners to validate methods, translate research into practical capabilities, and align Simulation Intelligence efforts with real-world constraints and product needs.

React-DeepONet: Developed a Deep Operator Network (DeepONet) based surrogate model for Stiff Chemical Kinetics.

• Employed a novel training mechanism and modified DeepONet architecture to learn the integration operator corresponding to large reaction systems.
• Added Param Net, Shift Net and learned Chemical Kinetics in latent space from Autoencoder for efficient training and robust predictions.
• Modeled and trained the modified DeepONet in JAX framework utilizing Just In Time compilation and Auto-Vectorization.
• Achieved stable and accurate long-time predictions with a huge Speed-Up of three orders (1000)

ChemNODE: Worked on development of efficient and robust Neural ODE models to learn stiff chemical kinetics.

• Implemented second-order optimizer (Levenberg–Marquardt) for efficient model training and robust predictions.
• Deployed Physics-Informed training for physically compliant and robust predictions.
• Latent Space Kinetics Identification through Neural ODE for large and complex fuels.

Reduced- Order Modeling of Turbulent Combustion (In collaboration with Sandia National Laboratories).

• Low- dimensional manifold is identified in form of Principal Components (PC) and these modes are evolved in time instead of thermochemical species in a direct numerical simulation.
• Chemical source terms and transport coefficients for the PC modes are mapped through high-fidelity Neural Networks.
• Achieved Speed-Up of two orders (80) with great accuracy on a laboratory-scale Bunsen Flame.

Input-Output analysis of Turbulent boundaries through Resolvent Analysis.

• Performed Resolvent analysis on the time-averaged turbulent mean flows.
• Identified the most responsive inputs and most receptive outputs of the dynamical system.
• Application to Shock- wave turbulent Boundary Layer Interaction (SBLI): Captured low-frequency unsteadiness and identified the most-energetic turbulent eddies after flow separation.
• Application to Compressible Turbulent Boundary layer: Obtained the dominant coherent structures in the boundary layer and confirmed their adherence to the attached eddy hypothesis.

Analysis of Thermal Buckling in a Jet Engine Compressor Disk.

• Investigated critical temperature for various modes using pre-existing literature on buckling of a solid disk.
• Modeled a solid disk and obtained the solution using eigenbuckling analysis on Ansys, matched with theoretical results.
• Predicted factor of safety for the Jet engine compressor disk using solutions procedure as used for the solid disk.

Computational analysis of suppression of the vortex shedding around a square cylinder.

• Numerically investigated flow around a square cylinder; discretizing Navier-Stokes equation by Marker and Cell Method.
• Identified suppression region and obtained the drag coefficient for various sizes of control cylinders and Reynolds numbers.