👨💼 I'm a PhD student in physics-based deep learning in the group of Prof. Thuerey at the Technical University of Munich (TUM), funded by Munich Center of Machine Learning (MCML). 🎓 I hold a master's with honours in Computational Science and Engineering from TUM. 💡 My areas of interest are scientific machine learning, differentiable physics, and AI4Science in general. 🌊 As a mechanical engineer, my first love is fluid dynamics. 💻 Previously I have worked as an Application Specialist for Cloud HPC Workflows at the Leibniz Supercomputing Centre in Munich.
This is my research at TUM that was published in ICLR 2025. We developed a method that reduces training cost for bilevel optimization scenarios, especially in physics-based deep learning. Our method speeded up the training of a neural-hybrid model for 2D Navier Stokes by 78% against the baseline.
2. Fluidchen
A finite differences solver for incompressible Navier-Stokes equations (in 2D) with heat transfer included. It contains own implementations of linear solvers: relaxation methods and Krylov subspace methods, is parallelized using MPI, and handles arbitrary geometries of the domain boundary as well as obstacles. This work was pursued in a team of 3 towards the Computational Fluid Dynamics Lab pratical course at TUM.
This is my contribution to the open-source coupling library preCICE. I took up this issue as a research assistant at the Chair of Scientific Computing at TUM. The PR is currently on hold due to changes in the software design and other recent decisions.
An intuitive web app that transcribes spoken input using whisper and gives summaries using the OpenAI API.
Feel free to reach out through:
- LinkedIn: kanishk-bhatia