Rajat Sarkar

About Me

I am an AI & ML Researcher at TCS Research, Pune and a PhD Scholar in the Department of Applied Mechanics at IIT Delhi, working under the supervision of Dr. Souvik Chakraborty. My doctoral research focuses on Hypergraph Computation for Scientific Machine Learning.

With a strong foundation in aerospace engineering and computational fluid dynamics, I bridge the gap between traditional physics-based simulations and modern deep learning. My work spans across developing novel graph and hypergraph neural network architectures for scientific computing, material imaging, battery life estimation, and industrial anomaly detection.

Research Interests

Scientific Machine Learning Hypergraph Neural Networks Graph Neural Networks Computational Fluid Dynamics Super-Resolution Computer Vision Physics-Informed ML Multiphysics

Education

2025 – Present

Ph.D., Applied Mechanics

Indian Institute of Technology (IIT) Delhi

CGPA: 10/10 · Advisor: Dr. Souvik Chakraborty

2018 – 2020

M.Tech, Aerodynamic & Flight Mechanics

Indian Institute of Space Science and Technology (IIST)

CGPA: 8.87/10

2013 – 2017

B.Tech, Aerospace Engineering

Hindustan University, Chennai

CGPA: 8.71/10

Experience

2020 – Present

Scientist

TCS Research, Pune

2019

Research Intern

Vikram Sarabhai Space Center (ISRO)

Gaganyaan Project · Boundary layer transition

Technical Skills

AI & ML

Graph Neural Networks Vision Transformers GANs Diffusion Models XGBoost

Programming

Python C MATLAB LaTeX

Software

OpenFOAM ANSYS Fluent Paraview PyTorch

Research

My research focuses on developing novel graph and hypergraph neural network architectures for scientific computing applications

Computers & Fluids 2026

Journal Paper

Implicit-PointSAGE

Neural Implicit Representation based Superresolution for Computational Fluid Dynamics

A mesh-independent super-resolution framework integrating PointSAGE with a Neural Implicit Learning module employing Galerkin-based attention. Efficiently captures underlying physics and predicts fine-mesh data directly from coarse-mesh inputs with 10-100X fewer training samples, delivering substantial computational acceleration for high-resolution CFD simulations.

Neural Implicit Galerkin Attention Point Clouds CFD

View Details

DMLR @ ICLR 2024

PointSAGE

Mesh-Independent Superresolution Approach to Fluid Flow Predictions

A novel mesh-independent framework leveraging point clouds to learn complex fluid flow and directly predict fine-resolution CFD simulations from coarse-resolution data. Combines Global Feature Extractor (PointNet-inspired) and Local Feature Extractor (GraphSAGE) achieving 30X-92X speedups over traditional CFD.

Point Clouds GraphSAGE CFD Super-Resolution

View Details

ML4PS @ NeurIPS 2023

Physics-Informed UNet (PIUNet)

Redefining Super-Resolution: Fine-mesh PDE Predictions without Classical Simulations

Proposes a novel definition of super-resolution for PDE-based problems using actual coarse-grid simulated data as input to predict fine-grid outcomes. The Physics-Informed UNet combines data loss with physics loss from governing equations, achieving up to 235X speedup over traditional CFD fine-mesh simulations.

Physics-Informed UNet PDE Combustion

View Details

PML4DC @ ICLR 2023

Vision HgNN

An Electron-Micrograph is Worth a Hypergraph of Hypernodes

A hypergraph-based backbone architecture for electron micrograph classification that models higher-order dependencies between image patches. Employs Hypergraph Structure Learning (HgSL), Hypergraph Attention Network (HgAT), and Hypergraph Transformer (HgT) modules for capturing short, moderate, and long-range spatial interactions.

Hypergraph Computer Vision Attention Micrographs

View Details

GLI @ NeurIPS 2022

Battery GraphNets

Relational Learning for Lithium-ion Batteries (LiBs) Life Estimation

A framework that jointly learns discrete dependency graph structures between battery parameters and models intrinsic degradation of lithium-ion batteries for Remaining Useful Life (RUL) prognosis. Features a Dynamic Graph Inference (DGI) module and Grapher module combining GNN and RNN blocks.

GNN Battery RUL Dynamic Graphs

View Details

ML4Materials @ KDD 2022

EMCNet

Graph-Nets for Electron Micrographs Classification

An end-to-end electron micrograph representation learning framework for nanomaterial identification. Represents images as patch-attributed grid graphs with three parallel pathways: Graph Encoder (GEnc), Hierarchical Graph Encoder (HGEnc), and Clique Tree Encoder (CTEnc).

Graph Networks Materials Science Classification

View Details

Hypergraph Learning for Anomaly Detection

Under Review

HgAD

Hypergraph Learning based Recommender System for Anomaly Detection, Control & Optimization

A self-adapting anomaly detection framework that jointly learns discrete hypergraph structures and models temporal trends among interdependent sensors. Features attention-based hypergraph convolution, hierarchical encoder-decoder, and root cause analysis with optimal predictive control.

Hypergraph Anomaly Detection Time Series Control

View Details

Publications

Peer-reviewed papers in top journals and ML venues

2026

Implicit-PointSAGE: Neural implicit representation based superresolution for computational fluid dynamics

Rajat Kumar Sarkar, Vishal Jadhav, Venkataramana Runkana

Computers and Fluids, Vol. 305, 106900 (Elsevier)

DOI

2024

PointSAGE: Mesh-independent superresolution approach to fluid flow predictions

Rajat Sarkar, Krishna Sai Sudhir Aripirala, Vishal Jadhav, Sagar Srinivas Sakhinana, Venkataramana Runkana

Data-Centric Machine Learning Research (DMLR) Workshop at ICLR 2024

2023

Redefining Super-Resolution: Fine-mesh PDE predictions without classical simulations

Rajat Sarkar, Ritam Mujumdar, Vishal Jadhav, Sagar Srinivas Sakhinana, Venkataramana Runkana

Machine Learning and the Physical Sciences Workshop at NeurIPS 2023

arXiv:2311.09740

2023

Multi-Knowledge Fusion Network for Time Series Representation Learning

Sagar Srinivas Sakhinana, Shivam Gupta, Sudhir Aripirala, Rajat Sarkar, Venkataramana Runkana

Machine Learning for IoT Workshop at ICLR 2023

2023

Vision-HgNN: An Electron Micrograph is Worth a Hypergraph of Hypernodes

Sagar Srinivas Sakhinana, Rajat Sarkar, Sreeja Gangasani, Venkataramana Runkana

Practical Machine Learning for Developing Countries (PML4DC) Workshop at ICLR 2023

2022

Battery GraphNets: Relational Learning for Lithium-ion Batteries (LiBs) Life Estimation

Sagar Srinivas Sakhinana, Rajat Sarkar, Venkataramana Runkana

Graph Learning for Industrial Applications Workshop at NeurIPS 2022

2022

EMCNet: Graph-Nets for Electron Micrographs Classification

Sagar Srinivas Sakhinana, Rajat Sarkar, Venkataramana Runkana

Machine Learning for Materials Workshop at ACM SIGKDD 2022

Patents

Intellectual property and inventions

High-Resolution Simulation Prediction For Computational Fluid Dynamics

Indian Patent App. 202421037547 · May 2024

Rajat Sarkar, Krishna Sai Sudhir Aripirala, Vishal Jadhav, Sagar Srinivas Sakhinana, Venkataramana Runkana

System and Method for Generating Mixed Variable Type Multivariate Temporal Synthetic Data

US Patent App. 17/994,580 · Dec 2023

Sagar Srinivas Sakhinana, Rajat Sarkar, Venkataramana Runkana

Privacy Preserving Generative Mechanism for Industrial Time-Series Data Disclosure

US Patent App. 17/804,682 · May 2023

Sagar Srinivas Sakhinana, Rajat Sarkar, Venkataramana Runkana

Congeniality-Preserving GANs for Imputing Low-Dimensional Multivariate Time-Series Data

US Patent App. 17/815,316 · May 2023

Sagar Srinivas Sakhinana, Rajat Sarkar, Venkataramana Runkana

About Me

Research Interests

Education

Ph.D., Applied Mechanics

M.Tech, Aerodynamic & Flight Mechanics

B.Tech, Aerospace Engineering

Experience

Scientist

Research Intern

Technical Skills

AI & ML

Programming

Software

Research

Implicit-PointSAGE

PointSAGE

Physics-Informed UNet (PIUNet)

Vision HgNN

Battery GraphNets

EMCNet

HgAD

Publications

Implicit-PointSAGE: Neural implicit representation based superresolution for computational fluid dynamics

PointSAGE: Mesh-independent superresolution approach to fluid flow predictions

Redefining Super-Resolution: Fine-mesh PDE predictions without classical simulations

Multi-Knowledge Fusion Network for Time Series Representation Learning

Vision-HgNN: An Electron Micrograph is Worth a Hypergraph of Hypernodes

Battery GraphNets: Relational Learning for Lithium-ion Batteries (LiBs) Life Estimation

EMCNet: Graph-Nets for Electron Micrographs Classification

Patents

High-Resolution Simulation Prediction For Computational Fluid Dynamics

System and Method for Generating Mixed Variable Type Multivariate Temporal Synthetic Data

Privacy Preserving Generative Mechanism for Industrial Time-Series Data Disclosure

Congeniality-Preserving GANs for Imputing Low-Dimensional Multivariate Time-Series Data

GATE AIR 71

ISRO Intern

Lunar Rover Finalist

10/10 CGPA

Get in Touch

Email

Phone

Location

LinkedIn